Orbit copies pstack (pstack/ in cursor/plugins) at commit 0b46ec3, the SHA that every skills.json entry pins. Every pstack. line quotes 0b46ec3, not newer upstream. The previous pin, e1cc031, differs from 0b46ec3 under pstack/ by three commits. e58dd04 moved the Gemini panel slug from gemini-4-carbon-high to gemini-4-neon-high, added setup-pstack's paragraph that migrates older Gemini entries, and reworded the Gemini cap. dd24b67 moved the slug to gemini-4-neon-xhigh and rewrote that paragraph and the cap to match. 0b46ec3 changed only statechart, which orbit does not vendor. So a line that says what e1cc031 holds or changed still holds at 0b46ec3. A line that names de8ee15, the pin before e1cc031, quotes text that orbit kept after e1cc031 cut or rewrote it. For Claude Code tool and path mappings, orbit follows pstack's adapter, pstack/docs/harness/claude-code.md. For T3 Code, orbit adds its own adapter, plugin/docs/harness/t3-code.md with t3-seats.md beside it (T3 Code host). On a design fork, check pstack first. Record each difference here with its reason.
In the orbit. lines, paths under agents/, hooks/ and skills/, rules.md and rules-shared.md are relative to plugin/, the folder the plugin ships from. Other paths are relative to the repo root.
- pstack. The user runs
/poteto-modeper task, typed as/pstack:poteto-modein Claude Code (README.md). - orbit. Four SessionStart hooks print the rules every session (
hooks/hooks.json). Three printrules.md, split at# Delegationand# Writing the reply, and the fourth printsrules-shared.md(Shared rules). The first line ofrules.mdsends a goal with no named skill through/orbit:start, and a go-ahead on a change Claude proposed outside a flow counts as a goal. Poteto-mode's Non-negotiables follow it.agents/poteto-agent.mdreadsrules-shared.mditself. - Why. Rules hold without the user remembering a command. No per-task command loads, so routing has to live in the rules. Claude Code caps each hook's stdout at 10,000 characters and replaces longer output with a file path and a preview of the first 2,000 characters (code.claude.com/docs/en/hooks, read 2026-11-17), so one hook would drop most of
rules.md. Two hooks split at# Code Qualityleft the second part at 9,931 characters at 0.55.0, so the next rules edit would have crossed the cap with no error. The four parts print about 6,300, 8,100, 7,300 and 1,800 characters at 0.71.0, andtests/hooks/session-start.test.shfails a part of 9,000 characters or more. The parts arrive in the order the hooks finish. Subagents don't get the SessionStart output. A "go" after a question-and-answer exchange once edited the main checkout with no worktree and collided with another session (transcript audit, 2026-11-10), so a go-ahead now enters a flow and its Worktree gate.
- pstack. poteto-agent reads all of poteto-mode's
SKILL.md, the lead's sections included (agents/poteto-agent.md). pstack has no rules file for agents alone. - orbit.
rules-shared.mdholds the four sections the lead and its delegates share. They are Bash, which is the "One plain command per Bash call" paragraph from rules.md Delegation, then Code Quality, Files and Comments. The text moved verbatim except for one sentence, and the# Bashheading is new. Comments opened with "Comments follow the same rule as the reply.", which pointed at Writing the reply in rules.md. It now names those rules itself, as short declarative sentences with no long-dash character and no colon joining two clauses.rules.mdkeeps the lead's sections. A fourth SessionStart hook printsrules-shared.md, so the lead's injected text holds the same sections as before (Loading rules). Therulesskill reads both files when the hook output is missing.agents/poteto-agent.mdreads the playbookSKILL.mdandrules-shared.md, and skipsrules.md. It carries one sentence of its own from Task End's Always-pause list, to stop and report to its parent before an irreversible write. The Delegate gate gives a prose unit's brief the paths ofskills/unslop/SKILL.mdandskills/technical-writing/SKILL.md. An autopilot owner's brief namesrules-shared.mdbesiderules.md(skills/playbook/playbooks/autopilot-full.mdstep 2). - Why. Across 1,474 subagents since 2026-11-09, 403 of 418 poteto-agent delegates read all of rules.md before working, because their agent file said to (transcript audit, 2026-11-20). rules.md was 3,498 words, about 7.7k tokens, and most of it is the lead's routing, triggers, delegation, shipping, Task End, reply style and Workflow Defaults. A delegate now reads 2,488 words before work in place of 5,660, the playbook's 2,183 and
rules-shared.md's 305. The lead's injected rules went from 3,498 words to 3,507, the new heading and the longer Comments sentence (measured withwc -w, 2026-11-20). A fourth hook, rather than the shared file appended to one part, keeps every part under the cap and lets the test compare the parts with the two files in order. Skippingrules.mddrops three things a delegate still needs, so each comes back in a narrower place. The Always-pause list lives in Task End, so poteto-agent carries its one sentence. Theunslopandtechnical-writingtriggers live in Non-negotiables, so the Delegate gate names both skills for prose units only. An autopilot owner is ageneral-purposeagent that runs a flow of its own and gets no SessionStart output, so its brief names both rules files (diff review of #91, 2026-11-20).
- pstack. pstack ships two agents,
poteto-agentandcomment-sicko, and neither carries a rule on the form of a shell command (agents/). - orbit. gitops, advisor, verifier and comment-sicko each carry the rule in their own file, in two to four sentences. Each forbids an
&&chain around git and a shell variable or$(...)in a git argument. gitops, the advisor and comment-sicko forbidgit -Corcdinto another checkout only while the session is isolated in a worktree, asrules-shared.mdBash does. Otherwise gitops and the advisor usegit -C <path>for a checkout their brief names. The verifier always runs isolated in its own worktree, so its clause stays unconditional and names another checkout of this repo, which leaves its scratch repo under/tmpallowed. gitops also names a heredoc and passes a commit body as a second-m. - Why. Subagents never get the SessionStart output, and these four agents read no rules file. In a ship cleanup eval run on 2026-11-19 (
tests/ship/cleanup.mjs), the advisor ranR=<root>; ... git -C $R config ...and gitops ran[ "$(git rev-parse HEAD)" = <sha> ] && git push origin HEAD:main. The guard refused both, with "names git in a form too complex to verify" and "redirects git to the shared checkout via -C". Across all sessions since 2026-11-09 the guard refused 687 calls, 1.75 h in all (transcript audit, 2026-11-20). The rule sits inline, not as a read ofrules-shared.md. gitops spawns several times per ship, and a read costs 305 words plus one more turn that re-sends the whole context, while the inline rule adds 98 words to gitops, 83 to the advisor, 51 to comment-sicko and 23 to the verifier (measured withwc -w, 2026-11-20). Thegit -Cclause is conditional because the T3 adapter (docs/harness/t3-code.mdWhere a child runs step 10) and autopilot-full step 2 send gitops into an owner's worktree withgit -C <path>from a session that is not isolated, and an unconditional ban would refuse that round (diff review of #91, 2026-11-20).
- pstack. poteto-mode's
SKILL.mdholds Non-negotiables, Principles, Autonomy, Subagents, Writing the reply and the playbook index. It has no gates. de8ee15 also had a Comments section. pstack ships its owncorrect. - orbit.
skills/start,skills/ship,skills/correct,skills/rules,skills/setupandskills/playbook/SKILL.mdare orbit's text, with noskills.jsonentry (Setup, Running another skill). The playbookSKILL.mdholds the gates, Principles, Stack rules and the index. poteto-mode's other sections live inrules.md.correctkeeps pstack's order (architecture, then a type or lint, then a test, then a doc) and its rule table, but handles one named mistake with one guard. It drops the class sweep, the exception rule and the Reply:. - Why.
startandshipare entry commands that pstack's single/poteto-modedoesn't need. The Worktree, Plan and Delegate gates are orbit's.rules.mdloads every session, so the always-on sections go there.correct's rewrite has no recorded reason.
- pstack. Orbit's
startentry point has no upstream counterpart (Orbit-written skills). - orbit.
skills/start/SKILL.mduses Claude'sTaskCreate/TaskUpdate/TaskListorTodoWrite. It falls back to a list in the reply only when no native task tool is available and the work is a multistep plan, meaning the goal holds two or more separate tasks the user would name. Single-task work, including building and landing one change, gets a line of progress instead. That list holds only the user's tasks, each with a checkbox. It shows once at the start, a task's line shows again when its state changes, and the final reply shows the closed list. The playbook's steps and gates are tracked silently and never printed. Steps update as they start and finish, skipped steps retain their reason and complete, and a resumed task reuses its list. - Why. The previous instruction named only
TodoWrite, so Codex and Claude sessions offering newer task tools fell back to text. Native updates let T3 display the flow's steps and progress in its task UI (operator request, 2026-11-17). A text list opened once and never updated, or closed with a sentence of evidence per item, hid progress. A reply list of playbook steps and gates was clutter, because the user needs only the tasks and their progress (operator request, 2026-11-19). A checklist for single-task work was noise, so it shows only for a multistep plan (operator request, 2026-11-19).
- pstack. e1cc031 cuts poteto-mode's Non-negotiables to ten short lines and adds one. "Starting any agent, whether a Task subagent or a Project worker → the first line of its kickoff is
Read <path>/worker.md before anything else."worker.mdholds the operator's rules for every started agent, among them "The bar in your kickoff is the whole bar. Add no check, gate, sign-off or review round above it." (poteto-mode/SKILL.md,poteto-mode/worker.md) - orbit. rules.md Non-negotiables keep the de8ee15 trigger list with orbit's edits, including "Nontrivial change, architecture decision, or "are we sure?" →
/orbit:how." Orbit ships noworker.mdand adds no kickoff line. Agents get the operator's rules throughrules-shared.md, their own agent files and the Delegate gate brief. - Why. rules.md is orbit's counterpart of poteto-mode's triggers, and it carries orbit and T3 Code triggers that the short list lacks. The prod-ready program rewrites rules.md in a later unit (P5), which can port the shorter wording.
worker.md's "Add no check, gate, sign-off or review round above it" would cut orbit's Plan gate and pre-land Independent verifier (P13 review, 2026-11-17).
- pstack. "Work from a git worktree off trunk, and keep unrelated work out of it" (
playbooks/opening-a-pr.md), with no tool named and no recovery steps. de8ee15 also had "Dirty branch with unrelated work: patch out, fresh worktree, apply. Snarled worktree: reset from main, redo minimally." - orbit. The Worktree gate (
skills/playbook/SKILL.md) callsEnterWorktreenamed<type>-<slug>, then runsgit branch -m <type>/<slug>. It carries both de8ee15 recovery sentences, which e1cc031 cut, and they override its skip-when-in-a-worktree rule. "Reset from main" becomes a fresh worktree. Either case runsExitWorktreewithkeepfirst. Shipping step 7 removes a task's worktree and branch once it lands, and/orbit:worktree-cleanupprunes abandoned ones (Shipping). - Why. Claude Code's worktree tools. gitops never runs
reset --hard(agents/gitops.md).EnterWorktreenames the branchworktree-<name>, so the gate renames it. Opening a PR pushes<type>/<slug>, and the worktree-cleanup audit matches PRs by local branch name.
- pstack. No setup step.
- orbit. Before entering, if
git check-ignore -q .claude/worktreesfails, the Worktree gate appends.claude/worktreesto.git/info/exclude. After entering, it runsvp installwhen the project has a lockfile. Gitignored files the Check needs, such as.env*or.claude/settings.local.json, come in through the project's.worktreeinclude, which Claude Code copies into each new worktree. When one is missing, the gate adds its pattern there and copies the file in once. - Why. Orbit's choice.
EnterWorktreeputs worktrees under.claude/worktrees, so the exclude keeps them out of the main checkout'sgit statuswithout a tracked.gitignoreedit. A fresh worktree has no installed dependencies, so the Check line can't run without the install. It also has no gitignored files, and agents copied.env.localin by hand in 4 sessions (transcript audit, 2026-11-10)..worktreeincludemakes the copy happen once per project instead of once per worktree.
- pstack. No plan gate and no advisor. pstack ships only
poteto-agentandcomment-sicko(agents/). - orbit. The Plan gate (
skills/playbook/SKILL.md) runs when a change adds a module or touches more than 3 files. The lead reads the project sources for those paths, briefsorbit:advisor(agents/advisor.md, Fable athighunless the project'sadvisorrow changes it, read-only) with the plan and the sources, and acts on its verdict. If the spawn fails because Fable is unavailable or its usage limit is reached, the lead retries once on Opus (rules.md Workflow Defaults, Models). The advisor has no T3 Code tools, because it only reads and returns a verdict. The advisor also picks a fork before it reaches the user (Test before asking) and reviews each PR's diff in Shipping (Diff review). - Why. Orbit's choice. A second model checks a multi-file plan against the project's rules before code exists, when a wrong shape costs least to fix. The advisor moved from Opus to Fable on 2026-11-10. Anthropic's Fable 5.2 launch page (https://www.anthropic.com/claude-fable-and-mythos-5-2) names Fable for the advisor role, where "Haiku, Sonnet, or Opus models consult it". The same page recommends Opus as the daily driver, so under the defaults the lead runs on Opus and a Fable advisor is a second model.
- pstack. "
Read<pstack>/skills/<name>/SKILL.mdand follow it. TheSkilltool refuses most pstack skills." (docs/harness/claude-code.md) - orbit. The Skill tool runs
start,playbook,ship,how,why,architect,arena,interrogate,swarm,figure-it-out,deslop,no-comments,unslop,technical-writing,control-ui,control-cli,pause-safely,worktree-cleanupandrules. Flows read the principle leaves,tdd,benchmark-checklist,show-me-your-work,ketchupandtypescript-best-practicesby path.rules.mdgivesbenchmark-checklistandshow-me-your-workthe bundled path,<orbit-plugin>/skills/<name>/SKILL.md. It givesbugbot-triage.mdthe bundled path<orbit-plugin>/skills/playbook/references/bugbot-triage.md. Skill bodies give a bundled path relative to the skill's own directory, such as../arena/SKILL.md.setup,correct,recall,reflect,automate-me,session-pickup,create-verification-skillandmaintain-verification-skillkeepdisable-model-invocation, so the user types them. - Why.
${CLAUDE_SKILL_DIR}expands only in a skill that the Skill tool loads or the user types.playbookandworktree-cleanupfind their files through it. Flows also pass$ARGUMENTS, which a Read by path can't. The SessionStart hooks printrules.mdandrules-shared.mdthroughsed, which rewrites<orbit-plugin>to the plugin root. Grok loads the plugin but drops the SessionStart output and leaves${CLAUDE_PLUGIN_ROOT}unexpanded in skill bodies. It gives the model each skill's base directory, so a relative path resolves there, and a lead without the hook output loads therulesskill, which reads../../rules.mdand../../rules-shared.md. Agents keep${CLAUDE_PLUGIN_ROOT}, which Claude Code and Grok expand in agent bodies.
- pstack. pstack's Claude Code adapter (
pstack/docs/harness/claude-code.md) maps Cursor names to Claude Code. Its Cursor-only features section skips what Claude Code lacks. "Claude Code has no Cursor Automations, Bugbot, Cursor cloud agents, or built-in PR tool. When a step needs one, put one line in your reply,Skipped in Claude Code: <step>. <reason>., then go on to the next step." - orbit.
plugin/docs/harness/t3-code.mdmaps pstack's names to T3 Code in three columns, pstack, T3 Code and T3 tool absent. It sits underdocs/harness/, the folder pstack uses for a host adapter. Before calling a T3 tool, each agent checks that tool itself. The tool is present when it is in the agent's tool list, or whenToolSearchwithselect:mcp__t3-code__<name>loads it. With noToolSearch, a tool not in the list is absent. T3 tools may load late, so the agent tries a missing tool's name once, byToolSearchselect:or else by its first real call, before it calls the tool absent. The probe has no families and never keys on another tool of the same family. Anymcp__t3-code__tool in the list tells the agent the host is T3, never that another tool is present. A tool that is not present takes the absent column. The rows mapTask,environment: "cloud", Cursor Automations,SendToUsersecret requests, a built-in PR tool, a wait on a PR event and background work the agent waits on. The Automations row adds that a request only to schedule callsschedule_taskand does not run its prompt now. The secrets row adds the sudo flow, adelivery: "file"secret fed tosudo -S -p '' <cmd>on stdin, with any other input as a file argument. Seven rows are orbit's own, since pstack's adapter has none of them. They are/loop, a question about a T3 thread by its id, a browser step the user asked for, a lead in T3 plan mode, gitops's CI wait, a verdict's proof table and the user's Stop./loopmaps toschedule_taskbound to the thread when that tool is present. A question about a thread runst3_thread_readon it before the answer, and without that tool the agent says it cannot read the thread and writes no skip line. A verdict's proof table goes throughhtml_previewwith an HTML file path or inline HTML, thenhtml_render, when both are present, else into an in-chat table, as swarm Phase D's result table. A chart for a trend or comparison, such as ketchup's, takes the same row, else a Mermaid chart (ketchup). The built-in PR tool row links each PR gitops creates, takes over or pushes to, withlink_pull_requestsfor several, and says when linking fails. The PR wait is two rows. gitops asked to wait on CI runs the watcher once with--onceon every host and reports the verdict (gitops agent). The lead's own wait follows babysit's Host watcher (Babysit). In T3 it reads once with--once, callswatch_pull_requeston a pendingTIMEOUT, ends the turn, and unwatches before a handback. In T3 with that tool off it reads once per turn and hands back. Outside T3 it runs the bare watcher under/loop. Bugbot and Cursor cloud agents keep the adapter's rows in both columns. Paths is one sentence, because T3 changes no path. Models keeps rules.md's Claude mapping and adds a row for a non-Claude slug, a panel seat on another family, with a Seats on another family subsection that points toplugin/docs/harness/t3-seats.md(Seats on another family). The lead callsorchestrator_capabilitiesonce per thread, and again only after an error or a user-named provider change, or withsearchfor a model the Models table names. The session's family is thedriverKindof the provider whoseproviderInstanceIdequalsinheritedProviderInstanceId. The roster is the list of entries in theother-familyrow's Model cell. An empty cell, no row andinheritgive no roster. When no provider matchesinheritedProviderInstanceId, the reply says so. A list then skips the seat with no entry tried, and with no list the session-family seat runs. The session-family seat is the seat the skill names, run on the session's own family as anAgentspawn. Theother-familyrow picks the seat (Model per role), and its Effort maps onto the model's own option id. The cell resolves to one of four outcomes, whicht3-seats.mdholds as one table (Ranked providers). An empty cell runs the session-family seat, and the reply says that no seat on another family is configured. A list tries at most three candidates outside the session's family, in list order, and never a provider it leaves out. Each candidate's model passes a family check on itsid, andautoanddefaultseats sayfamily unverified. When every candidate fails, the seat is skipped and the reply says so, unless the list ends withinherit. Before the lead calls an explicit pick unavailable because its model is not listed, it callsorchestrator_capabilitieswithsearchset to the model id. The lead spawns the seat withdelegate_task, ends the turn, reads the result withtask_statuswhen the child's completion wakes it, and settles the child thread. Each spawn carries aclientRequestIdnaming the skill, the seat and the round, reused on a retry of the same call and new for a new round or a fallback. A seat or child that runs commands skips a provider whose listedruntimeModeslacks the parent's runtime mode. A child's question goes throught3_pending_requeston itschildThreadId, answered from the brief's context or else by the user, andarchiveis only for a child the lead abandons or when that tool is absent. A failed seat puts<entry> unavailable: <reason>in the reply and moves to the next candidate, aftertask_cancelon a child that still runs.t3_thread_sendonly answers or steers a running child. In the absent column the session-family seat runs, and the reply names each seat's model. "Roster" has no pstack name. pstack says "family", which arena Phase A step 3 defines by slug prefix, and "pool", as inarena cross-judge pool. Read-only and shared-checkout children stay foregroundAgentcalls. Under a T3 lead withdelegate_task, a child that needs its own checkout or runs while the lead keeps working, such as a parallel delegate, the native verifier or an autopilot owner's build round, is a same-providerdelegate_taskchild, and a panel seat on another family is a cross-provider one (Own-checkout children). EachAgentcall passesrun_in_background: false, and parallel spawns go as several calls in one message (Agent call defaults). A new round's brief carries the original brief, the prior round's findings, the responses to them and the objections still open. A skipped step gets the lineSkipped outside T3: <step>. <reason>.The built-in PR tool row's absent column isgh pr createand writes no line, as pstack's adapter says, "That is not a skip, so write no line." rules.md Delegation points at the file once, in Host tools. The one exception to the probe is a session connected to T3 from outside. When a call there returns the error that the tool needs an agent running inside T3 Code, that tool and the rest of the tools that act as the calling thread take the absent column, andt3_thread_launchnever stands in for them unless the user asked for a new thread. A verifier whoseverifierrow readsother-familyorother-family, inheritruns as a seat on another family throughdelegate_task(Independent verifier). The native verifier runs as a same-provider child (Own-checkout children). A browser step the user asked for runspreview_statusfirst, thenpreview_openonly if no tab can be automated, and closes a tab it opened witht3_preview_closebefore the reply, unless the user asked for it or uses it. No agent usespreview_*unasked to verify or show its own edit. Without the tools, the step runs on the control-ui harness. A lead in T3 plan mode reads its mode fromorchestrator_capabilitiesand passesinteractionMode: "plan"on everydelegate_task, never"default", and omits the field when unsure. A native child's brief carries T3's plan-mode rule verbatim, and the lead spawns no code writer in plan mode. The user's Stop ends the thread's PR watches and stops its delegated tasks. Nothing re-arms on its own, and resumed PR work goes back through the lead's PR wait row (Playbooks as skills). The read trigger in the adapter's opening names the browser step and a child in plan mode. - Why. T3 Code has a tool for three of the adapter's rows.
schedule_taskcovers Automations,request_secretcovers secret requests, andlink_pull_requestfollowsgh pr createfor the built-in PR tool. T3's runtime rule for every thread (runtimeInstructions.tsline 3 in pingdotgg/t3code at c2eb8f3) says "Link each PR you create, take over, or push to: link_pull_request, or link_pull_requests for several, full URL, every stack layer, as your next call. Host tools don't link for you. If linking fails, say so." The old row's "A later push to a linked PR needs no new link." contradicted "push to", andlink_pull_requestis idempotent (Tool.Idempotentin T3'spr-tools.ts), so a repeat link costs nothing. The same line sets the PR wait, quoted under Babysit. Bugbot and cloud agents have no T3 tool, so their rows match the adapter's. A probe beats an environment variable, because T3 can turn a tool off inside a T3 session and the variable says nothing about which (T3-native plan, rejected alternatives, 2026-11-16). The probe checks each tool, not one per family, because T3 can turn one tool off and leave the rest of its family on (inferred). pstack keeps its host adapter atdocs/harness/claude-code.md, so orbit's T3 adapter takes the same folder underplugin/. The T3-native plan renders each verdict's proof table in the coordinator thread withhtml_render, and swarm Phase D already returns its result table in chat, which covers a session without the tool. Being in T3 also doesn't give every agent the tools. Anorbit:verifiersaw none until itstoolsfrontmatter named them (T3 probe lane, 2026-11-16), so each agent probes for itself. Owners, verifiers and swarm workers once ran asdelegate_taskchild threads. The threads cluttered the sidebar, needed settling by hand, and died in a T3 restart as a subagent would (T3-native plan, rejected alternatives). So orbit moved them toAgent, which gives a worktree, parallel runs and fresh context. T3's rule sends a separate checkout and a background agent todelegate_task, so P7 of the prod-ready program moved the children that need either back to child threads and kept the read-only ones native (Own-checkout children). pstack ships a Claude Code adapter beside its Cursor-native text, so the absent column keeps orbit working in plain Claude Code and in a T3 session with a tool turned off. The skip line says "outside T3" because an absent tool can also mean a T3 session with that tool off (inferred). The PR row once had the lead writeSkipped outside T3: link PR., and underclaude -pwith no T3 the reply carried it on every PR (interrogate, 2026-11-16). The probe lane on 2026-11-16 also saw/loopwith no interval schedule its wake withScheduleWakeup, as the/looprow says. pstack's adapter maps a non-Claude slug as "Omitmodelso the agent runs on the session model, and say so." Orbit's absent column departs from that line. Each skill's Claude seat already is the session-family seat orbit runs without T3, and a caveat on every non-T3 run is noise (advisor plan gate, 2026-11-16). Orbit keys family ondriverKindbecause the capabilities result carries no family field, anddriverKindis its only per-provider kind (inferred). A provider of anotherdriverKindcan still list a Claude model, and some list a model nameddefault, so an explicitother-familypick is the fix (advisor, 2026-11-16).t3_background_waitrefusesdelegate_taskids ("Subagent, Agent tool, and delegate_task IDs don't go here.", T3-3 probe, 2026-11-16), so the lead ends the turn and the child's completion wakes it. An aggregator provider of anotherdriverKindcan host a model of the session's own family (advisor, 2026-11-16), so the family check reads each candidate model'sid(Ranked providers).orchestrator_capabilitiesreturns a short list and takessearch, so a model missing from the first result can still be in the catalog. The skill's Claude seat became the session-family seat when other providers could lead, and the name stayed. T3's orchestrator MCP doc (c2eb8f3) says a call withoutclientRequestIdcreates new work, a provider runs a runtime mode it does not offer as approval-required, approvals are not routed to the parent, and a child's question waits ont3_pending_request. The same doc'sdelegate_tasksection says "Each independent delegated review round uses a newdelegate_taskcall with the original brief, prior findings, responses, and unresolved objections." (2026-11-16, re-checked 2026-11-17) T3's system prompt for every thread (orchestrationInstructions.tsin pingdotgg/t3code at c2eb8f3) sets six more rules. The probe retry is its "T3 tools may load late. Try the name once before you call it absent." (line 20). The capabilities rule is its "Callorchestrator_capabilitiesonce per thread, before the firstdelegate_task. Again only after an error or a user-named provider change." (line 10). Orbit keeps thesearchcall for a named model beside it, because the first result is a short list. The thread row is its "Given a thread id,t3_thread_readit before you answer about it." (line 15). That prompt carries the rule from the thread's first turn, so the row needs no trigger in rules.md Host tools. The sudo flow and the schedule-only rule are its "Sudo: ifsudo -n <cmd>needs a password, usedelivery: "file"and runsudo -S -p '' <cmd> < "$file"; the password takes stdin, so give other input as a file argument. Asked only to schedule:schedule_taskwithout running its prompt." (line 19). The proof table follows its "pass an HTML file path or inline HTML tohtml_preview, thenhtml_render" (line 23). T3's outside-agents doc (docs/user/outside-agents.mdin pingdotgg/t3code at c2eb8f3, read 2026-11-16, re-checked 2026-11-17) says the tools that act as the calling thread, "subagents, secret requests, rendered HTML, the browser preview, devices, and worktree handoff", "still appear in an outside agent's tool list, but calling one returns an error saying it needs an agent running inside T3 Code." So for a session connected from outside, a listed tool is not proof, and that error marks the whole group absent. The same doc hands an outside agent's work off witht3_thread_launch, but T3's own rules start a thread only on request, so the probe never swaps it in for an absent tool (user directive, 2026-11-16). The browser row copies T3's system prompt. "Never usepreview_*unasked to verify or show your own edit. ... Close a tab you opened before you reply, unless the user asked for it or uses it." (orchestrationInstructions.ts:24) and "When browser work was asked,preview_*is the browser you share with the user:preview_statusfirst, thenpreview_openif no tab can be automated." and "Switch browsers only if the tools are absent orpreview_opensays unsupported, the user asks, or two preview calls fail on one step. Then quote the error and switch." (line 36). The plan-mode row follows T3's orchestrator MCP doc, "For local delegation, provider, model, runtime mode, and interaction mode inherit from the parent when omitted" (line 245) and "A child interaction mode may stay equal to or narrow fromdefaulttoplan. It may not escalate fromplantodefault." (line 534). A wrong explicitplanwould narrow a child the lead meant to run in default, so an unsure lead omits the field and T3 inherits it (inferred). A nativeAgentchild gets none of T3's mode text (inferred), so its brief carries T3's ACP plan-mode instruction (line 68). The Stop row followsdocs/orchestration-v2/feature-lifecycles.md:254, "A Stop also holds the queue, ends the thread's pull request watches, and drops wakes its delegated tasks still owe, so nothing automatic restarts the thread." T3's source-control doc adds that watching ends "when you press Stop on the thread" and that "Subagents cannot watch pull requests" (docs/user/source-control.md:340-345). The adapter is two files, where pstack has one per host. rules.md Host tools. has every T3 lead readt3-code.mdwhole, and at 6,144 words it held a section most sessions never use. Seats on another family applies only when theother-familyrow lists providers or theverifierrow isother-familyorother-family, inherit. It moved verbatim intot3-seats.md, andt3-code.mdkeeps a pointer that names the condition. A lead with no roster reads 3,429 words in place of 6,144, and one with a roster reads 5,365 (measured withwc -w, 2026-11-20, before 0.70.0 cut the adapter's lead-provider lines). The pieces of the seats section that hold for every child stayed int3-code.md, because Where a child runs uses them for every same-provider child. They are theorchestrator_capabilitiesparagraph, the Effort-to-target.optionsparagraph, theworkStaterule, thelinkrule, thet3_background_wait,t3_thread_interruptandarchiverules, and the rule for a child's pending question. The skills that use a roster seat condition on theother-familyrow listing providers, so a lead with no roster never openst3-seats.mdto find out. Where a child runs step 5 namesagents/<name>.mditself, so a verifier child needs no second file. The pointer also states the reply line for an emptyother-familyrow, which a lead with no roster still writes.
- pstack. pstack's text uses Cursor names, and its one adapter,
pstack/docs/harness/claude-code.md, maps them for Claude Code. It maps no other host's names. - orbit. Claude, through Claude Code, is the only lead. The
rulesskill tells a lead that is not running in Claude Code that orbit leads only under Claude Code, and stops. Codex, Grok and other families run only as panel seats on another family and as the verifier seat, throughdelegate_task(Seats on another family)./orbit:startuses only Claude's task tools, and/orbit:setupno longer links Codex or checks what Codex and Grok load (Setup). - Why. The user decided on 2026-11-20 that leads on providers other than Claude made orbit hard to manage. Rules carried a fork on the lead's provider, with its own tool names, model defaults, load path and reading rules. Seats on other families stay, because a second family still checks a Claude author's work. Before 0.70.0, Codex and Grok could lead a T3 thread through a second adapter file, and
docs/trails/keeps that history.
- pstack. No actor named. "Commit liberally" (
playbooks/feature.mdstep 4). "Rebase into small, ordered commits" (feature.mdstep 6) and "Rebase into small, landable commits" (opening-a-pr.md). Bodies follow/technical-writing. - orbit. The parent reviews each unit, then
orbit:gitopscommits it by path, with a subject and a body that says why, undertechnical-writing(agents/gitops.md). Unpushed fixes fold in with fixup and autosquash. Feature step 6 and refactoring step 8 keep pstack's rebase into small, ordered commits. gitops runs it before Shipping, with fixup and autosquash on unpushed units, and never rewrites a pushed commit. Refactor commits land in step 8's order, and bug's failing repro is its own unit before the fix. - Why. Review comes before commit, and git output stays out of the main context. Orbit once dropped the rebase step on the grounds that units committed in order leave nothing to rebase. Fix rounds and parallel units still leave commits worth folding, and the user asked to "club all 3 commits into one so i can see overall changes" (transcript audit, 2026-11-10). So the step matches pstack again. A pushed commit stays, because rewriting it takes a force-push.
- pstack. No git agent. The working agent commits and runs the forge.
- orbit.
agents/gitops.mdruns every commit, push, rebase and forge write, withmodel: sonnetandeffort: low. Agitopsrow in the project's Models table can change either (Model per role). It never deletes a branch, except in shipping step 7, where it removes the worktrees still left for a landed branch by absolute path from its starting directory, never throughcdor-Cand never with--force, and deletes the branch. It never switches, merges or fast-forwards that starting checkout, which is the user's. In T3 the lead removes each settled child's worktree first, as the adapter says (Shipping). Its tools addToolSearchand three named T3 tools,link_pull_request,link_pull_requestsandlist_thread_pull_requests. Before linking a PR it checks for the link tools per the adapter's probe (T3 Code host), and that check gates only the link. When they are present, gitops links each PR it creates, takes over or pushes to, as its next call, withlink_pull_requestsfor several. It skips PRs it only reads or reviews and says when linking fails. Asked to wait on CI, on every host, it runs orbit'swatch-pronce with--onceand reports the verdict and reason. After PR work its last line isPR <full URL>, head <SHA>, checks and reviews <state>. It never runsgh pr checks --watch. It waits this way even when the lead's brief names another command or a re-check loop. On a conflicted cherry-pick it runsgit cherry-pick --abortbefore it reports. - Why. Orbit-only agent, so no pstack model applies. Git and
ghwork is routine, so it runs on a cheaper tier than the session by default. The cleanup exception lets gitops run the step it would otherwise stop at. In the ship cleanup eval at d840cd4 (2026-11-19), gitops removed a worktree throughcd <root> && git worktree removeand the worktree guard refused it, while agit worktree remove <absolute path>with no redirect went through (Shipping). T3 tools are deferred, so gitops needsToolSearchto load them. In the T3 probe lane on 2026-11-16, gitops with this tool list loaded and calledlist_thread_pull_requests, and${CLAUDE_PLUGIN_ROOT}expanded in its body, so the watcher path resolves. Naming three PR tools keeps gitops off T3's thread, schedule and preview tools (inferred). The wait waswatch-pr --timeout 120, which sleeps between polls for up to 120 s. T3's runtime rule (runtimeInstructions.tsline 3 at c2eb8f3) says "Delegated children never watch, even if the brief says babysit." and "For PR waits, no sleep, loops, gh --watch, Monitor or ScheduleWakeup." So gitops makes one--onceread, which never sleeps and needs no Bash timeout. The last line is the same rule's "After PR work, a child's last line is: PR , head , checks and reviews ." The link rule follows the same line (T3 Code host). gitops reports once and returns, so the lead, which stays in the thread, owns any longer wait (inferred). In one T3 lane the lead namedgh pr checksand a periodic re-check in its brief, and gitops never probed and slept in a loop between checks (T3 lane, 2026-11-16). When the brief named no command, gitops probed and ranwatch-pronce. The old absent branch rangh pr checks --watchand ran it again when the watch timed out, which is the loop shape that lane caught. pstack has nogh pr checks --watch. Its babysit waits on GitHub CI with the watcher (playbooks/babysit.md), and a code search of cursor/plugins finds the command only incursor-team-kit, outside pstack (2026-11-16). So gitops waits the same way on every host, and rules.md Host tools. has the brief name the PR, not a command. A cherry-pick left in progress would block the lead's serial rerun of the unit on the task head, so gitops aborts it (P7 review, 2026-11-17).
- pstack. No such hook. pstack ships no hooks (
pstack/plugin.json). - orbit. A PreToolUse hook on Bash (
hooks/no-claude-trailer.sh) refuses agit commitorgit commit-treeat command position whose command, or the message file-F/--filenames, holds aCo-Authored-By: Claudetrailer, and tells the agent to commit again without it. A command that only mentions both strings passes. It never reads the description. - Why. Agents added the trailer despite an empty
attribution.commitand gitops's "No trailers". Leads copied it into briefs, and later agents copied it fromgit log, so the text rule didn't hold. A hook holds it structurally (principle Encode Lessons in Structure).
- pstack.
setup-pstackwrites~/.claude/rules/pstack-models.md(docs/harness/claude-code.md). Flows name a model slot each, such as "your configured feature model (defaultclaude-opus-5-5-xhigh-fast)" (playbooks/feature.md) and "your configured hillclimb model" (playbooks/hillclimb.md). The defaults are "Code, prose, and judgment:claude-opus-5-5-xhigh-fast" and "Trivial mechanical edits:claude-opus-5-5-high-fast" (poteto-mode/SKILL.mdSubagents). The adapter maps a Claude slug to its model and effort and says "Drop-fast". "Agemini-*or other non-Claude slug" maps to "Omitmodelso the agent runs on the session model, and say so." "Verification always runs on theswarm workersmodel,claude-opus-5-5-high-fastby default." The adapter's setup step says "Never put the role lines inCLAUDE.mdorAGENTS.md." de8ee15's defaults were "grok-4.7-xhigh-fastfor code,claude-opus-5-5-xhighfor prose and judgment", and the hardest changes went "to your strongest judgment model". - orbit. A
### Modelstable in the project's Workflow block replaces most ofpstack-models.md, with| Task | Use for | Model | Effort |rows (rules.md Workflow Defaults, Models). The user file~/.claude/rules/orbit-models.mdholds rows in the same format, and/orbit:setupwrites itsother-familyandverifierrows (Setup). Each cell comes from the project's table, else from the user file, else the default. The per-flow slots become per-unit work types that the project names, such asshader-mathordata-tuning. The Delegate gate runs each code delegate on the work type whose Use for fits its unit, at that row's model and exact effort. A work type's cell may add, <value> after <n> failed tries. A failed try is a delegate whose work fails review or the check, or that crashes or returns nothing. Parallel attempts count no tries, and each part of a split unit starts at zero. A unit that fits two work types splits when each part verifies on its own, else runs as the type of its hardest part, and row order breaks a tie. A unit that turns out to fit another type is re-briefed under it, and its try count starts over. With no work types, or none that fits, a prose delegate (docs, written explanations) runs on thejudgmentrole and code delegates tier by difficulty. The hardest run withmodel: "opus"(thejudgmentrole). Trivial mechanical edits run on thecoderole, and so does every other delegate unless the project'scoderow issonnetorhaiku. Then the rest run on the session model. The lead's reply names each unit no work type fit. Six Task names are orbit's roles.other-familyis a panel seat on another family under T3 Code (Seats on another family). Its empty Model runs the skill's own seat, and the reply says that none is configured.inheritkeeps the session-family seat. Otherwise its Model is an ordered list of<providerInstanceId>or<providerInstanceId>/<model>entries, which tries at most three listed providers outside the parent's family, in list order, as pstack'sarena cross-judge pooldoes.inheritmay end the list (Ranked providers). A bare<providerInstanceId>takes that provider's first eligible model.judgmentdefaults toopus. It is every spawn a skill names withmodel: "opus"or as thejudgmentmodel, such as interrogate's first adversary, arena's first runner, autopilot owners and thehowexplainer. A skill'smodel: "opus"names that role, so it is not a step's own model.codedefaults toinherit. It is interrogate's second adversary, arena's second runner, swarm workers,recall's source investigators and every spawn no role or work type covers.gitops,verifierandadvisordefault to their agent's frontmatter model and effort, and rules.md Delegation hands gitops its row's model and effort when the table has agitopsrow.orbit:comment-sickotakes no row and runs on its frontmatter, so acoderow doesn't move it. Panels name their roles, so they never take a work type. The lead passes the resolved model as theAgentcall'smodel, which overrides the agent's frontmatter, andinheritomits it, so the spawn gets Claude Code's default subagent model (CLAUDE_CODE_SUBAGENT_MODELwhen set, else the session model). A step that names its own model or effort keeps it, such asrecall'shaikuand the trail reviewer. rules.md's "Match the model to the job" passes an effort a step or row names as written, and otherwise picks one for the job, with no speed. The bug, feature, refactoring, perf and hillclimb playbooks point pstack's line at rules.md, as "Pick each subagent's reasoning level for the job, per the Delegation section of rules.md." Their delegate steps drop pstack'sconfigured <flow> model (default ...)slug for the Delegate gate.swarmandarenareword it to pick each worker's or runner'seffortper "Match the model to the job".interrogateruns arena's step 3 and sets each adversary'seffortto high (Review).howandshow-me-your-workdrop it, andwhyandreflectspawn nothing. The verifier takes theverifierrow, not thecodemodel, and keeps its owneffort: highwhen the row names none.agents/poteto-agent.mdcarrieseffort: mediumandagents/comment-sicko.mdcarrieseffort: medium, which pstack's agent files don't. The lead still picks poteto-agent's level, so its frontmatter only catches a missed pick. Theverifierrow's Model may also beother-familyorother-family, inherit. Under T3 Code with a roster, the lead then fills each PR's verifier as a seat on another family, picked as theother-familyrow picks its seat. When no candidate is left, it runs the native verifier only if theverifierModel isother-family, inherit, and otherwise asks the user or holds the PR (Independent verifier)./orbit:setupwrites that row asother-family, inheritorother-family, from the user's answer to its second question. With no such row, the verifier keeps its frontmatter. No other row takes a provider list, so theother-familyrow stays the one cross-provider list. A spawn that fails because its model is unavailable or at its usage limit retries once on the session model, and anorbit:advisoron Fable retries on Opus. No native spawn moves to another provider. - Why. At de8ee15 the adapter's setup defaults mapped to
opus xhighand toinherit-parentfor every grok role, which is orbit'sjudgmentdefault ofopusandcodedefault ofinherit. pstack's 17 role lines then held only those two values by default, so two roles,judgmentandcode, carry every default, and the prose and judgment lines tookjudgment. e1cc031 cut the role lines to nine, moved every default to an Opus 5.5-fastslug at xhigh or high, and gave the panels Fable and Gemini seats. Orbit drops that routing and keeps its two roles (Effort ceiling, Panel seats). A table in the project's AGENTS.md travels with the repo and sets the mapping where the project's cost and quality tradeoff lives. The adapter keeps one user's model picks out of the repo (inferred), and orbit's table holds the project's own work types, which belong in the repo. Which T3 providers a user has enabled is the user's, not the project's, so theother-familychoice also lives in a user file, as pstack's model rule does, and a project row still wins. Per-unit work types are the user's call, 2026-11-07.Taskis the user's word for a row's name. "Work type" names the project's rows, so they don't clash with the playbook's task types (bug,feature) or with orbit's roles.inheritis Claude Code's frontmatter value for the session model. Claude Code has no Auto model, so pstack'sinherit-parentandautobecome one value. Theafter <n> failed triesstep and the re-brief came with the work types (5a3d474), with no recorded reason. A unit moves to a costlier model or effort only after the cheaper one has failed on it, and a re-briefed unit runs under its new row's cells, so tries under the old row don't count (inferred). comment-sicko has no line insetup-pstack, so it gets no row. One flow mixes kinds of work, so a bug fix can needshader-mathfor its fix anddocsfor a doc unit, and a per-flow slot can't tell them apart. The verifier takes its own row because the user named it as its own role, so a cheapercoderow doesn't move it. pstack sends "Trivial mechanical edits" to "your fast code model" and leaves the middle tier to each playbook's line. Orbit has no playbook lines, so a cheapercoderow once took every delegate short of the hardest. The user said "never code with sonnet unless its very small" (transcript audit, 2026-11-10), so asonnetorhaikucoderow now gets only the trivial edits. The guard first compared thecoderow's price with the session model's. In a Fable session an Opuscoderow counts as cheaper, so that test sent every delegate short of the hardest back to Fable. Namingsonnetandhaikukeeps the guard on the user's reason and lets a Fable lead's code run on Opus (Fable lead). Claude Code has no per-spawn speed, so the fast-mode rule is dropped. The routing eval (tests/routing/run.mjs) caught leads that passed no effort, gitops in 2 of 3 runs and comment-sicko once, so every orbit agent's frontmatter now holds a level the spawn falls back to. poteto-agent's level moved fromhightomediumon 2026-11-11. The user said "in most cases: for coding opus 5.5 medium is good enough" and "allow poteto-agent to run high if the task requires it". A missed pick on a hard unit now runs at medium, which the user accepts.other-familyis its own role, not thecoderow moved to the roster. The T3-native plan's Appendix B rejects that move, becausecodealso covers code delegates, swarm workers and live lanes, and a child on another provider loads no orbit rules. The name follows pstack's "another family". The operator asked on 2026-11-16 to default to the current agent for most work, unless more providers add real value, and named the verifier as the one helper that prefers another family. pstack keeps verification on theswarm workersmodel. That was a grok slug by default at de8ee15, so a Claude lead's verifier ran off its family there too. e1cc031 moved the default toclaude-opus-5-5-high-fast. The advisor plan gate (2026-11-16) ruled out per-row fallback lists. Their,separator collides with, <value> after <n> failed tries, setup cannot know the lead's family, and a project cell overrides the whole user cell. The same verdict kept native fallbacks on the lead's provider, because moving gitops or a half-finished code child to another provider loses its hooks, itstoolslimit and the worktree guard. The retry on an unavailable model departs from T3's system prompt, "Provider or model unavailable: say so. Never switch unasked or fall back to native tools, Workflow or CLIs." (orchestrationInstructions.ts:10in pingdotgg/t3code at c2eb8f3). It switches the model unasked. Orbit keeps it only for nativeAgentspawns, keeps each spawn on its provider, and has the reply name the model that ran, so the switch is said. A seat on another family never retries this way. It moves only to the next entry the user listed, and otherwise skips or asks (Ranked providers). Without the retry, a Models row whose model is at its usage limit would stop every spawn on that row until the user edits the row (inferred).
- pstack. No provider ranking and no cap. The
arena cross-judge poolline lists slugs in order, and "The judge is never from the parent's vendor". When a slug fails, arena uses "the nearest valid slug of the same family, then the next such entry, or skip the judge, and say so." (skills/arena/SKILL.md) - orbit. The host adapter resolves the
other-familyModel cell to one of four outcomes in one table. An empty cell or no row runs the session-family seat, and the reply says that no seat on another family is configured and that/orbit:setuplists providers.inheritruns the session-family seat. A list gives at most three candidates, one perproviderInstanceId, in list order. A candidate's provider is outside the session'sdriverKind, is enabled, can run a cross-provider child task, lists models and, for a seat that runs commands, offers the parent's runtime mode. Its model passes the family check, which runs before the listing check. The check splits the model'sidat-,/,.and:, and refuses a token that equals the session's family word or names its vendor,anthropicforclaude,openaiforgptandxaiforgrok. That word isinheritedModel's, else, for an alias such asopus, that of the session provider's first model whoseidholds the alias as a token. With neither, the session's family is unknown, no entry fails the check on that ground, and every seat saysfamily unverified.autoanddefaultare opaque. A bare entry takes one only when every model its provider lists is opaque, and that seat saysfamily unverified. The lead tries the candidates in order for one seat and never tries a fourth. When every candidate fails, the seat is skipped with the lineSkipped <skill>'s seat on another family: every listed entry failed., and only a list that ends withinheritruns the session-family seat. The verifier never skips. With noinheritin theverifierrow, the lead asks the user before a native verifier runs, or holds the PR when no user can answer. Setup's question about thatinheritrecommends no./orbit:setuprecommends the first three enabled providers ranked Codex (codex), Muse (muse), Gemini (gemini), then Grok (grok), and byproviderInstanceIdwithin a kind. Setup names a typed entry that falls after the third outside the session's family, says that a seat tries it only when an earlier entry fails the seat's checks, and accepts the list. The ranking lives only in that recommendation. - Why. T3's system prompt for every thread says "Provider or model unavailable: say so. Never switch unasked or fall back to native tools, Workflow or CLIs." (
orchestrationInstructions.ts:10inpingdotgg/t3codeatc2eb8f3). Before this change, an empty cell filled the seat from a provider the user never listed, and a failed list fell back to the session's family without asking. A provider the user listed counts as asked (prod-ready plan, Appendix B row 2), so the list is the roster, and a trailinginheritis the user's written yes to the native fallback. Setup first recommended yes to its native verifier question, on the advisor's call, so that a PR would not wait on the user. The operator replaced that default with no, the strict reading of "Never switch unasked or fall back to native tools" (operator, 2026-11-17). With no, a PR whose listed verifiers all fail holds and the lead asks. A skipped seat follows pstack's arena, which skips its judge when no entry is left. The same prompt says "Every new delegate review or re-check is a newdelegate_task(brief, findings, objections), never at3_thread_send." (line 11), sot3_thread_sendonly answers or steers a running child. Issue #27 on aayushbtw/orbit found thatdriverKindis not the model's family. An aggregator of anotherdriverKindcan serve the session's own family, and a model nameddefaultsays nothing about what it runs, so the check reads the model'sid. Issue #28 asked setup and the runtime cap to agree, so both count the first three entries outside the lead's kind. The runtime counts candidates after each seat's checks, so a later entry runs when an earlier one fails them, and setup names that entry instead of refusing it (adversarial review, 2026-11-17). The session's family word once came frominheritedModeland from every model the session's own provider lists. An aggregator lead lists several families, so that rule refused every seat (adversarial review, 2026-11-17). The operator asked to rank providers, keep at most three, and named the order Claude, Codex, Muse, Gemini, Grok (2026-11-17). Claude left the ranking when it became the only lead (user decision, 2026-11-20). A liveorchestrator_capabilitiesresult on 2026-11-17 listed Muse withdriverKindmuse, disabled, and no Gemini provider, so thegeminiid is unverified. In the same result fx, kiro and pi listed onlydefault. A model id such asopenai/o3has no family word, so a gpt-family session once passed it through the family check although OpenAI serves it. Mapping three vendor names to their family words was a one-clause change, smaller than a newfamily unverifiedcase for every vendor prefix, and it agrees with the lane grader (P9 review, 2026-11-17). Live T3 lanes under a Claude lead on claude-opus-5-5 ran four cases (tests/t3/fixtures/).seat-emptyandseat-all-failran at 41b1c74, andseat-listedandverifier-all-failran at 31ee0b3. Untested live: theinheritoutcome, a list that ends withinherit, the family check against an aggregator or a vendor prefix, the cap, the verifier's hold with no user, and setup's ranked recommendation and second question.
- pstack.
setup-pstackdetects the model slugs aTaskcan take, loadspstack-models.mdcwhen it exists, asks for a budget, shows every role and asks to accept or change roles, validates each slug against the detected set, overwrites the whole file "so re-runs stay idempotent", and tells the user it "applies to new sessions" (skills/setup-pstack/SKILL.md). The Claude Code adapter writes~/.claude/rules/pstack-models.mdwith no frontmatter, because "Claude Code loads every.mdfile in~/.claude/rules/into each new session", and says "Never put the role lines inCLAUDE.mdorAGENTS.md." Steps 7 and 8 offer verification skills. - orbit.
skills/setup(/orbit:setup), orbit's text with noskills.jsonentry. Setup runs on any host, and only its T3 provider step keeps setup-pstack's shape. Its first step runsskills/setup/scripts/doctor.sh, a read-only script that writes no files, needs no network and always exits 0. The doctor checks git, gh and its login, node, bun, jq, the git repo andorigin/main, the Workflow block and install drift. setup-pstack has no such check. Each change that setup could make becomes an offer. Setup asks about every offer in one multi-select question at the end and changes nothing without a yes. It drops the budget, the role list and the verification-skill offers.docs/setup.md, a doc an agent followed to install orbit per provider, is gone. When the install record's version differs from the version of the folder Claude Code installs orbit from, setup offersclaude plugin marketplace updateand thenclaude plugin update. That folder is the marketplace'sinstallLocationjoined with the plugin'ssource. With no install record, as under--plugin-dir, setup skips the drift check. In a git repo with no Workflow block and novpr, setup offers a filled block fromskills/setup/workflow-block.mdfor the root AGENTS.md, neverCLAUDE.md. The lead finds the project's Check command in the rootpackage.jsonscripts orMakefileand never runs it. The template moved there fromtemplates/workflow-block.md, which now only points to it. The T3 provider step runs only whenorchestrator_capabilitiesis present per the adapter's Probe. Otherwise setup prints a skip line and goes on. The step keeps setup-pstack's detect, load current state, show and confirm, validate, write and confirm. It shows every provider, enabled with its models or disabled, and marks the session's family. It asks two questions in one call, with no budget and no other role. The first asks which providers in order fill theother-familyseat, and recommends the ranked list from Ranked providers as fallbacks for one seat, not several reviewers. The second asks whether orbit may run its native verifier when no listed provider can run it, and recommends no (Ranked providers). Validate checks each entry against the capabilities result, withsearchfor a named model. It refuses a repeated provider and aninheritbefore the last entry, keeps a trailinginheritit found, and accepts an entry that falls after the third outside the session's family. It names that entry as tried only when an earlier entry fails a seat's checks. The write goes to~/.claude/rules/orbit-models.mdwith no frontmatter, one header line and a Models table in rules.md's format. setup-pstack overwrites its whole file. Orbit's write replaces only the Models table and keeps every line outside it byte for byte. It also writes averifierrow when the provider list is not empty, with Modelother-family, inheritafter a yes to the second question andother-familyafter a no. An existingother-familyrow reads ascurrent: no native fallbackwith the question unanswered, because setup before 0.60.0 wrote that row without asking. - Why. The user's goal, 2026-11-16, is one install through Claude Code's
/pluginlines, then one command in T3 that saves which providers orbit uses, user-wide, with no per-provider installs. Claude Code has no per-spawn speed or budget knob, and the user caps effort at high, so the budget question has nothing to set. Roles other thanother-familyalready live in the project's table (Model per role). Orbit's other opt-in skills keepdisable-model-invocation, and setup writes a user file, so the user types it. Setup once linked~/.codex/skills/orbitto orbit's folder, checked what Grok loaded and offered an always-on rules line for each, so either could lead. Those steps left with the other leads (Claude is the only lead). Setup once needed T3 Code and stopped without it, so a plain Claude Code user had no way to find a missing tool before a flow failed on it. The doctor runs on every host and checks the tools that orbit's flows call (production-readiness program, P11, 2026-11-17). A verifier saw a lead runbun run checkin the user's project before any yes, so setup reads the files that name the Check command and never runs it (verifier, 2026-11-17). Onlyplugin/ships, so the template moved into the setup skill, where setup can read it on an installed copy (AGENTS.md). A table-only write is as idempotent as a whole-file overwrite and keeps the user's own lines (inferred).
- pstack.
setup-pstackasks for a budget,large,mediumorsmall. "largekeeps" each default's effort, "mediumlowers it one step, andsmalllowers it two steps", and "No role runs abovexhigh" (skills/setup-pstack/SKILL.md). poteto-mode's Subagents section defaults "Code, prose, and judgment" toclaude-opus-5-5-xhigh-fastand panels to "claude-fable-5-2-high,claude-opus-5-5-high-fast, andgemini-4-neon-xhigh". "Raise any other role one step, up to xhigh, for the hardest or subtlest work, lower it one step for judgment, and two or more steps for routine work." "Opus 5.5 always runs its-fastslug." "Fable 5.2 is an exception. It runs outside a panel only when the user named it for this task or two Opus 5.5 xhigh runs of the same brief failed its check." If the tool rejects a slug, "never pass it again", fall back through the role's line or default, and "Say so." (poteto-mode/SKILL.mdSubagents) - orbit. No budget setting and no starting level to step from. rules.md's "Match the model to the job" has the lead pick high for hard and judgment work, medium for a code writer even on a small change, low only for an edit that writes no logic (a typo fix or one fixed line), and medium or low for other routine work. It passes xhigh or max only when a step, a Models row or the user names it. A spawn that runs on Fable gets medium or high, never low. The verifier's frontmatter runs at high. If the
Agenttool rejects a level, the lead tries no other level, uses the row's level or the agent's default, and says so, as pstack does. rules.md Delegation keeps the de8ee15 Subagents text with orbit's edits. It names no-fastslug, no xhigh default and no limit on Fable. Its Writing the reply section stays whole, plus one link bullet (File links) and less the bullet that Final reply cut. Comments moved torules-shared.md(Shared rules). - Why. The user found xhigh not worth its tokens, 2026-11-07, and said it again on 2026-11-11, "do not use xhigh use max high when u need it". pstack's ladder starts every role at xhigh, so orbit names the levels directly instead of steps from it. The Models table already lets a project name xhigh where it pays off, so a budget knob adds nothing. The Fable floor is the user's rule, "dont use low fable" (2026-11-11). e1cc031's Opus-first routing (poteto-mode hunk 2) starts code, prose and judgment at xhigh, which breaks that ceiling. It also runs code on Opus 5.5 at xhigh, where the user said "in most cases: for coding opus 5.5 medium is good enough" (2026-11-11). Its Fable exception would bar the advisor, which runs on Fable at every plan gate (Plan gate). The same hunk cut Writing the reply to one paragraph and removed Comments, which orbit's rules.md keeps, so the hunk stays out whole (P13 review, 2026-11-17). "Medium or low for routine work" let Claude leads pass low to every code delegate in the live matrix at 34fdaae (code-child, foreground and parallel-units lanes), so a code writer now gets medium per the user's line. At cfc882d the Codex parallel-units lead still passed low to both implementation children, because it read a small new module (
min.mjsormax.mjswith a node:test test) as a trivial mechanical edit. So low now covers only an edit that writes no logic, such as the typo fix in autopilot-owners or the one fixed line in code-child and foreground (#71).
- pstack. Writing the reply names no link form. "Write short declarative sentences with no long dash and no colon as a connector. Lead with what changes for the user. Give each claim its evidence or a label (measured, inferred, or guess). Never hand the human a check you could run." (
poteto-mode/SKILL.md) - orbit. rules.md Writing the reply adds one bullet after "Never fabricate", "Embed images, video and audio with Markdown image syntax and an absolute path. Link every other file you cite or produce by its absolute path, as in
[http.ts:42](/abs/path/http.ts:42)." It holds on every host. - Why. T3 asks for that form in every thread. Its runtime prompt says "You can embed images, video, and audio using Markdown image syntax with absolute file paths. Link every other file you cite or produce by its absolute path, as in http.ts:42, so the user can open it in T3 Code's files viewer." (
runtimeInstructions.ts:23in pingdotgg/t3code at c2eb8f3). rules.md holds one rule for every host, so a plain Claude Code reply carries the same links.
- pstack. Writing the reply is one paragraph. "Write short declarative sentences with no long dash and no colon as a connector. Lead with what changes for the user. Give each claim its evidence or a label (measured, inferred, or guess). Never hand the human a check you could run." (
poteto-mode/SKILL.md) pstack has no rule that keeps every section a playbook's reply names, and no fixed shape for the final reply. Each playbook's Reply: lists what its reply holds. - orbit. rules.md Task End holds a Final reply paragraph. The final reply has four parts, in order. They are the outcome in one or two sentences, the closed task list for a multistep plan, an always-present
Heads-up:list, and theNext:line. The outcome restates the goal, says whether it is met, and gives the PR link or the landed commit. Checks that passed get one clause there.Heads-up:lists, worst first, a part of the goal not done, a skipped or waived step, a check that failed, did not run, passed with parts unverified or ran on a fallback, an unasked behavior change or accepted tradeoff, anything left behind, a problem outside the goal or a dismissed finding, and a reversible call made in the user's place. With none it readsHeads-up: nothing skipped, failed or left behind.What ran as planned stays out of the reply, and under Landprthe PR body holds it. When no PR opens, what a playbook or Non-negotiables sends to the PR body goes between the outcome andHeads-up:. A Reply: item that is the result itself, such as an answer, a number, a path, a plan or a status per PR, goes there too. Any other Reply: item that ran as planned goes in the PR body, or in that slot when no PR opens. Writing the reply drops orbit's "Terse is not an excuse to drop content." bullet, which kept every section a playbook's reply named. Its last sentence now says each playbook's Reply: names what the task reports, and Final reply says where each part goes. The principles that shaped a decision go in the PR body's## Tradeoffs, or in the reply when no PR opens. opening-a-pr.md andagents/gitops.mdwiden## Tradeoffsfrom only rejected alternatives to also hold those principles./orbit:startstep 3 gives each task list line only the task and its state. The feature, bug and shipping Reply: lines say which item goes in the outcome, which inHeads-up:, and which in the PR body. Bug's repro output goes in a PR comment or file that## Verificationlinks, per opening-a-pr's log rule. Shipping lists each verdict other than a plainPASSwith its notes.docs/harness/t3-seats.mdandt3-code.mdname a seat's model only when it fell back. - Why. The final reply joined the feature, Opening a PR and Shipping Reply: lists, and the Terse bullet forbade cutting any of them. On a PR of UI fixes the final reply ran about 30 lines, with a goal restatement, a detailed checklist, "How it was checked", "Process notes" and "Cleanup". The user called it noise and said "it should show important stuff: what it didnt do, anything important, info that user should know, nothing should fail silently" (user, 2026-11-20). The
Heads-up:line is always present, so a skip, a failure or a fallback never drops out silently. The PR body already holds the detail for a reviewer.
- pstack.
model: claude-opus-5-5-xhigh-fastin thepoteto-agentfrontmatter (agents/poteto-agent.md). - orbit.
agents/poteto-agent.mdhas nomodelkey and setseffort: medium. Each delegate runs on the model its work type or role resolves (rules.md Workflow Defaults, Models), at the effort the lead picks per "Match the model to the job". - Why. The user caps effort at high (Effort ceiling), and an xhigh
-fastslug in the frontmatter would win over a missed pick. Orbit's Models roles, not one fixed slug, choose each delegate's model.
- pstack. "
run_in_background: true, agent mode (readonly strips MCP), file pointers not inlined context" (poteto-mode/SKILL.mdSubagents). - orbit. rules.md Delegation keeps agent mode and file pointers. Agent mode is
general-purposeor anorbit:agent, notExplore. pstack'srun_in_background: truedrops out. When anymcp__t3-code__tool is in the tool list, everyAgentcall passesrun_in_background: false, and parallel spawns go as several calls in one message. On other hosts the call passes norun_in_background. The T3 adapter'sTaskrow says the same (T3 Code host). Models follow Model per role. - Why. The adapter maps
readonly: falsetogeneral-purpose, which keeps the session's MCP servers, andreadonly: truetoExplore. It omitsrun_in_background: truebecause subagents run in the background by default. T3 forbids that default. Its system prompt says "Subagents: use your native tool when the child uses your provider, a model it offers and no settings beyond yours (ClaudeAgent,run_in_background:false; Codexspawn_agent, then wait). Usedelegate_taskfor another provider's model, user-asked settings, a separate checkout (isolation:"worktree"), ordelegate_taskby name. A native type on another provider's model counts too. Background agent:delegate_task, never a backgroundAgent. Workflow only when asked." (orchestrationInstructions.ts:9in pingdotgg/t3code at c2eb8f3). The rule sits in rules.md, not only in the adapter, because rules.md is in context before the adapter is read, so a T3 session's first spawn follows rules.md alone. The rule holds only in T3, because the lead waits for every call in a foreground batch before it can act again. In plain Claude Code that wait would stop autopilot-full's root from firing its/looptick, its code-ready swarm and a stuck-owner replacement while owners build. In T3, a child that needs its own checkout or runs while the lead keeps working is an asyncdelegate_taskchild, so rolling starts, the code-ready swarm and the tick hold there too (Own-checkout children).
- pstack. Children that need their own checkout or run beside the lead are Cursor cloud agents. Swarm spawns its workers "with
subagent_type: generalPurpose,environment: "cloud",run_in_background: true" (skills/swarm/SKILL.md). In autopilot-full, "Each owner is a Cursor cloud agent" (playbooks/autopilot-full.mdstep 2). Arena gives each candidate "a git worktree where possible" and spawns them all "withrun_in_background: true" (skills/arena/SKILL.md). The Claude Code adapter mapsenvironment: "cloud"toisolation: "worktree"(docs/harness/claude-code.md). - orbit. The T3 adapter's Where a child runs table (
plugin/docs/harness/t3-code.md) picks how each child runs, by what the child needs. Its three columns are a T3 lead withdelegate_task, T3 otherwise (the agent is a child or a subagent, ordelegate_taskis off), and not T3. A child that shares the lead's checkout is a foregroundAgentwith no worktree in all three columns. That row holds the read-only children, such as thehowexplainer, the advisor, judges on the session's family, audit and verify lanes, and swarm workers that only read, plus gitops, serial delegates, small review fixes, comment-sicko, and an owner's fix, rebase and merge rounds, which work in the owner's worktree. A read-only lane reads a SHA throughgit showor agit archive -oexport, never through a checkout of its own. A child that needs its own checkout is a parallel delegate, an arena runner, an interrogate adversary, a swarm worker that writes, or the native verifier. Under a T3 lead withdelegate_task, it is a same-providerdelegate_taskchild. In T3 otherwise, its units run one at a time in the agent's own checkout, never as anAgentwithisolation: "worktree"and never in a worktree made in the shell. Outside T3, Claude Code runs it as anAgentwithisolation: "worktree", where the delegate commits its own unit and gitops cherry-picks it, as in T3. A child that runs while the lead keeps working is an autopilot owner's build round. Under a T3 lead withdelegate_task, it is a same-providerdelegate_taskchild too. In T3 otherwise, owners run one at a time in the foreground, and the reply says the tick cannot fire while one runs. Outside T3, owners stayAgentcalls withisolation: "worktree", rolling under the Parallel cap. Panel seats and the verifier seat on another family keep Seats on another family. No T3 lane callsAgentwithisolation: "worktree". The lead callsorchestrator_capabilitiesonce per thread before the firstdelegate_task, and probesdelegate_task,task_status,task_cancelandt3_thread_organize. Each call passes atargetwith the lead'sinheritedProviderInstanceId, the model the spawn's Models role resolves and its effort as the model's effort option, elseoptions: []. It also passesisolation: "worktree",mode: "async", arole(implementationfor writers and owners,reviewfor adversaries,testfor the verifier), atitlethat names the skill and the unit, and noruntimeMode. ItsclientRequestIdis<skill>-<unit>-<short base sha>-try<n>. A retry of the same call after a tool error or a lost response reuses the id, and a new try or round takes a new one. Before the spawn, the lead commits what the child needs on the task branch. Each brief names the start SHA, the task head, or for an owner or other work off trunk the SHA oforigin/<trunk>after a fetch, or for an owner stacked on another PR that PR's tip. It also names the lead's checkout and the project root. The child's first command isgit rev-parse --show-toplevel, and when it prints either path the child stops and reports without moving HEAD. Next it runsgit reset --hard <start sha>. A verifier or adversary checks out the SHAs its brief names instead. A brief to anorbit:agent names the agent file,agents/<name>.md, and the plugin folder, and states the agent'stoolslimit. Ageneral-purposechild, such as an owner, a swarm worker or an adversary, gets the plugin folder and the playbookSKILL.mdpath instead. Every brief says "You are a subagent of this thread. Start no child thread." An owner's build round is its only child. Its brief allows native foreground subagents and drops the babysit loop and the merge-ready stop. Under Landprthe owner pushes, opens its PR, reads its state once withwatch-pr --once, and ends withPR <url>, head <sha>, checks and reviews <state>. The owner never watches a PR, and the root runs babysit on it indrivemode. The child commits its unit by path in its own worktree and reports itspwd, branch and commit SHA. The lead neither waits nor polls, works on while other work is left, and ends the turn when none is left. The child's completion brings the wake. With a turn running, the wake arrives inside that turn, beside the next tool result. With no turn running, the wake is a new run after the turn ends. A wake that arrives in the middle of a step, such as a review or another child's cherry-pick, waits until that step is done, and the lead handles each wake once. On the wake, the lead readstask_status. AworkStateother thanresult_availablemeans the child still runs, so the lead neither reads nor settles it and waits for its wake. Atresult_availablethe lead checksgit merge-base --is-ancestor <start sha> <sha>, and reviewsgit show <sha>from its own checkout. A commit that does not descend from the start SHA is a failed try, never cherry-picked. gitops runsgit cherry-pick <sha>in the task worktree. On a conflict, gitops runsgit cherry-pick --abortand reports, and the lead reruns the unit serially on the new task head as a nativeorbit:poteto-agent. Then the lead settles the child witht3_thread_organizeand runsgit worktree list. The lead, not gitops, removes a child worktree that remains withgit worktree remove <path>, and its branch withgit branch -D <branch>when its tip is the picked commit. A fix round after the cherry-pick is a nativeorbit:poteto-agentin the task worktree. Only a full redo of the unit is a newdelegate_task, with a new id, and it counts one failed try. A stuck child getstask_cancelbefore any replacement, and "run fewer" or "no delegate" cancels the extras and starts none. Until a child's wake, the tick judges it by its trail and its git side effects, never bytask_status. The lead settles an owner's build child and keeps its worktree. Each later round of the role, a fix, a rebase or the merge round, is a foregroundAgentin the lead's thread with noisolationand noEnterWorktree, briefed with the owner's worktree path and branch. It is anorbit:poteto-agenton thejudgmentmodel for code, ororbit:gitopsfor git and forge work. Only whengit worktree listno longer shows that worktree is the round a fresh child, whose brief runs the same guard, thengit switch -C <branch> <head sha>, and stops and reports, never forces, when another worktree holds the branch. Inside a child, units that need their own checkout run one at a time. - Why. T3's system prompt for every thread (
orchestrationInstructions.tsin pingdotgg/t3code at c2eb8f3, read 2026-11-17) sets five rules for children. Line 9 says "Subagents: use your native tool when the child uses your provider, a model it offers and no settings beyond yours (ClaudeAgent,run_in_background:false; Codexspawn_agent, then wait). Usedelegate_taskfor another provider's model, user-asked settings, a separate checkout (isolation:"worktree"), ordelegate_taskby name. A native type on another provider's model counts too. Background agent:delegate_task, never a backgroundAgent. Workflow only when asked." Line 10 says "Callorchestrator_capabilitiesonce per thread, before the firstdelegate_task. Again only after an error or a user-named provider change. Settings go intarget.options. Provider or model unavailable: say so. Never switch unasked or fall back to native tools, Workflow or CLIs. Never make a child's checkout in the shell or intask." Line 11 says "Every new delegate review or re-check is a newdelegate_task(brief, findings, objections), never at3_thread_send. Fix small review findings yourself; start no delegate for them. When told to proceed without a delegate, or to run fewer,task_cancelthe extras. Start no replacement." Line 12 says "Follow a running delegate'snext. Don't poll it (task_status,sleep, watchers, Monitor, ScheduleWakeup). Stop a delegate run you replace or no longer need withtask_cancel, nevert3_thread_interrupton achildThreadId." Line 13 says "As a subagent or child, do the work yourself. Start no subagents unless your task says to." AnAgentwithisolation: "worktree"is a separate checkout, and an owner that runs while the root works is a background agent, so line 9 sends both todelegate_task. Line 10 rules out a brief that has the child make its own worktree in the shell. Line 11 keeps small fixes native, so an owner's fix, rebase and merge rounds run in the lead's thread and start no delegate. Line 13 keeps a child from starting its own child thread. T3's runtime prompt (runtimeInstructions.tsin pingdotgg/t3code at c2eb8f3) adds "Delegated children never watch, even if the brief says babysit", "For PR waits, no sleep, loops" and "After PR work, a child's last line is: PR , head , checks and reviews ." So the build child reads its PR once and ends on that line, and the root, which stays in the thread, runs babysit. The operator wants as few threads as the rules allow, because each thread costs attention, so only a child that needs its own checkout or runs long becomes one (operator directive, 2026-11-17). The prod-ready program ran this way. Each owner's build round was adelegate_taskchild, and each fix, rebase and merge round ran in the root's thread as a foregroundorbit:poteto-agentororbit:gitopspointed at the owner's worktree (prod-ready program, 2026-11-17). T3's orchestrator MCP doc (docs/orchestration-v2/orchestrator-mcp-server.mdin pingdotgg/t3code at c2eb8f3, read 2026-11-17) says to "useisolation: "worktree"for a separate prepared checkout from the parent's commit", and "Uncommitted parent edits are not copied." So the lead commits first. The parent's commit there is HEAD of the parent thread's bound worktree, else of the project root, anddelegate_tasktakes no base ref (OrchestratorMcpService.tsin pingdotgg/t3code, about lines 2809-2836, read 2026-11-20).EnterWorktreedoes not move T3's binding. In threechild-base --rootlanes at 7686e86 (2026-11-20), all six children of a thread opened on the project root started at the root's commit, without the lead's committedlib/round.mjs(#88). The adapter had said "The child's checkout starts from your commit." Leads coped three ways. One reran its round after the children stopped, one brief happened to saygit merge --ff-only <sha>, and in one the children fast-forwarded on their own. So every brief now pins the start SHA. T3 makes a worktree only whenisolation: "worktree"is passed. Without it the child runs in the thread's bound checkout, which for a thread on the project root is the user's own checkout, where the reset would move the user's branch and drop their edits. So the child first checks that its checkout is neither the lead's nor the project root (diff review of #89, 2026-11-20). The child resets its own fresh branch there, becausegit merge --ff-onlyfails when the start is not behind the task head (inferred), and because a detached checkout would leave the branch at T3's start, while the lead deletes a child's branch only when its tip is the picked SHA. A child commit off another base would carry that base's history into the review, so it is a failed try. The same doc says "When the result includesnext, completion wakes the parent: do not wait on the task or poll it; continue other work, and end the turn when none is left." So the root starts the next owner as one finishes and fires its code-ready swarm and its tick while owners build, which a foreground batch blocked (Agent call defaults). Under a Claude lead, T3 used to deliver the wake only as a new run after the turn ended. pingdotgg/t3code#39528, merged 2026-11-20, changes onlyClaudeAdapterV2.tsand its test, and reports one notice in the running turn and no second run. It shipped in T3 nightly 0.1.11-nightly.20261120.4979. A live probe in a Claude lead on that build on 2026-11-20 started an asyncdelegate_taskchild on Haiku that repliedPROBE_CHILD_DONE. The lead ran oneRead, and the notice "Delegated task (probe mid-turn delivery) in child thread [completed]: PROBE_CHILD_DONE" arrived beside thatRead's result in the same turn.task_statusthen showedworkStateresult_availablewith the same summary, and the settle succeeded. The mid-step wait and the rule to handle each wake once are orbit's own reading, so a wake never cuts into a review or a cherry-pick and a handled child is never read again. Two cases are unknown. No run has shown whether a child's pending question arrives mid-turn. A wake while the lead is blocked in a foregroundAgentcall is inferred to land at the tool boundary when that call returns, and is unprobed. A wake can still find the task unsettled. In pingdotgg/t3code at 966e422 (read 2026-11-20), thetask_statusdescription inapps/server/src/mcp/toolkits/orchestrator/tools.tssays "Do not call it on a running task before its completion wakes you.", and the test "finalizes a child whose run ended before %s settled" inapps/server/src/orchestration-v2/DelegatedCompletionDelivery.test.ts(about line 5744) keeps the parent's task open after the child's run ends until the child's background command or nested subagent settles. So a state other thanresult_availablegets no read, settle or poll, and the lead waits for the wake. No live run has shown that state on a wake. The doc also saysclientRequestId"derives stable command, thread, and message IDs within the provider session. Retrying the same call returns the same durable work." The id carries the base SHA so that a second run of the same skill and unit in one session starts new work instead of returning the first run's (advisor plan gate, 2026-11-17). A prototype on 2026-11-17 ran a same-providerdelegate_taskchild withisolation: "worktree"onclaude-haiku-5-5from parent commit1956d10. The child'sgit rev-parse --git-common-dirwas the main checkout's.git, and its tree was clean, without the parent's uncommitted edits. The child committed024650f, and the parent'sgit cat-file -t 024650freturnedcommit, so the lead can cherry-pick it from the shared.git. After the settle, the worktree and the branch were gone (prod-ready plan, Appendix A). The advisor plan gate of 2026-11-17 still made the cleanup idempotent, withgit worktree listafter the settle and removal of what remains, so the step converges whether or not a settle removes the worktree (P7 trail). Outside T3 the Delegate gate once had gitops commit each kept unit inside the child'sisolation: "worktree"worktree. Claude Code 2.1.326's worktree guard refuses that write, " is isolated in the worktree , so its writes are limited to that folder. This path is in a different worktree ()." A headlessparallel-unitsrun at 34fdaae worked only because the lead had each child commit its own unit and gitops cherry-pick by SHA, as T3 does, so every host now runs that way (#70). The P7parallel-unitslane on 2026-11-17 needed it. After each settle, the lead'sgit worktree liststill listed the child's worktree, and the lead removed it and the child's branch. In the same lane gitops rangit cherry-pick <sha>in the task worktree for both units with no refusal (tests/t3/fixtures/claude-parallel-units.json). The child commits in its own working directory, and the lead reviews and cherry-picks by SHA from its own checkout, because Claude Code's worktree guard refusedgit -Cinto a delegate's worktree (docs/trails/pstack-harm-audit.tsv, 2026-11-10). A fix round runs native because line 11 says to fix small findings without a delegate. The prod-ready root did so for P2, with a nativeorbit:poteto-agentin the owner's worktree (prod-ready trail, 2026-11-17). The owners' worktrees forfeat/t3-seat-selection,feat/t3-pr-waitsandfeat/t3-delegate-childrenstayed under~/.t3/worktrees/orbit/after their build children returned (git worktree list, 2026-11-17), so a later round finds the worktree there. The prototype above saw a settle remove a unit child's worktree, so the lead checksgit worktree listbefore a later round and falls back to a fresh child only when the worktree is gone. Autopilot step 4 lanes read the code-ready SHA fromgit archive -o <scratch>/head.tar <sha>, unpacked withtar -xf. Orbit reads line 10 as a ban on a child's checkout, and an export is a folder of files, not a checkout, so a read-only lane needs no thread (orbit's reading, P7 plan, 2026-11-17). The shipping verifier checks out the head SHA and moves HEAD, so it is a child with its own checkout. Interrogate adversaries keep their repros in their own checkout, so in T3 each one is a child too, one thread per adversary (P7 plan, 2026-11-17). The prod-ready plan rejected keeping every child a nativeAgentwithisolation: "worktree", because "it breaks T3's rule that a separate checkout or background work goes throughdelegate_task" (plan Appendix B row 1). The root applied the advisor's verdict, owners asdelegate_task, because the operator said to use the advisor for decisions (prod-ready trail, 2026-11-17). The same row says "If the operator prefers fewer threads over T3 conformance, P7 drops out and the clash goes intodocs/pstack-differences.mdas a recorded divergence." An operator who asks for no threads reverts this default. The cost is threads. Under a T3 lead withdelegate_task, exactly these children become child threads in T3's sidebar: a parallel delegate, an arena runner, an interrogate adversary, a swarm worker that writes, the native verifier, an autopilot owner's build round, and a panel seat or the verifier seat on another family. Every other child is a foregroundAgentthat shows as a nested node. That covers an owner's fix, rebase and merge rounds, verify and audit lanes, workers that only read, judges on the session's family, the advisor, thehowexplainer, gitops, serial delegates and comment-sicko. The lead settles each child thread atresult_available.
- pstack. "Resume an existing subagent only when the work needs state that lives in it and is costly to move" (
poteto-mode/SKILL.mdSubagents). de8ee15 said "Resume, message, or queue a follow-up on an existing subagent only when the new work strictly needs state that lives in that agent and is costly to move". - orbit. The de8ee15 paragraph in rules.md Delegation, plus one sentence. It stays because poteto-mode hunk 2 stays out (Effort ceiling). A resume means a new agent with the earlier result pasted into its brief. A fresh agent that takes a returned owner's next round first enters the prior worktree (Autopilot-full). Under a T3 lead with
delegate_task, that round is a foregroundAgentin the lead's thread that works in the owner's worktree by its path, with noEnterWorktree(Own-checkout children). - Why. The adapter maps
resumeto spawning a new agent and pasting the earlier result into its brief. That agent starts in its parent's working directory, not the prior agent's worktree.
- pstack. No cap. Swarm says "Spawn all N workers in one message" with
environment: "cloud"(skills/swarm/SKILL.md), and autopilot-full says "Run owners in parallel off trunk" (playbooks/autopilot-full.mdstep 3). - orbit. rules.md Workflow Defaults item 5. At most 4 agents that build, test or run the app run at once, and read-only agents don't count. Swarm spawns up to the cap in one message and starts the next as one finishes. Autopilot owners, parallel delegate units and hillclimb, figure-it-out and arena fan-out run under it. Under a T3 lead with
delegate_task, they run as same-providerdelegate_taskchildren and still count (Own-checkout children). - Why. pstack's workers run on cloud VMs, so their count costs the user's machine nothing. Here every agent runs on the user's laptop. The user said "too many parallel agents i told u laptop is heating up bro" and "use 4-5 max" (transcript audit, 2026-11-10). The default takes the low end of that range. A configurable
Parallel:key and rules for counting nested agents were tried and dropped on 2026-11-10, because pstack has no such machinery. A project that needs fewer agents says so in its own AGENTS.md.
- pstack. Workers run on cloud VMs. Swarm uses
environment: "local""only when the worker needs access to something on the user's computer" (skills/swarm/SKILL.md). - orbit. Every Delegate gate brief and every swarm brief carries one line verbatim, "You run on the user's machine, not a VM. Run the app headless or windowless when the project allows it, and never open a window that takes focus or plays sound unless the user agreed."
agents/verifier.mdcarries the same line. - Why. pstack's prose assumes a VM, where a window or a sound reaches no one. Here a launched app takes the user's screen. The user said "stop spawning fullscreen game instances its shifting my screen" (transcript audit, 2026-11-10).
- pstack. Arena's defaults name three runners, Fable, Opus and Gemini, and a judge "never from the parent's vendor" (
skills/arena/SKILL.md).interrogate reviewersdefaults to the same three slugs (skills/setup-pstack/SKILL.md). - orbit. When the Models table resolves
codeandjudgmentto the same model, as an Opus session does under the defaults, both interrogate adversaries and both arena runners run on that model. Arena then skips the judge, because no model other than the parent's is left (Arena judge). Panels name their roles, so a work type never changes their model. Under T3 Code with a roster, arena's judge and the other seats on another family run on the roster, so an Opus session keeps a second family unless every listed entry fails (Seats on another family). Without one, the text above holds. - Why. This follows the adapter's grok mapping (Model per role). Without a roster, the panel loses its model diversity in that case, and orbit accepts that to stay on the adapter. A project that wants diversity there adds a
coderow with a different model.
- pstack. No Fable lead. Its defaults name
claude-opus-5-5-xhigh-fastfor code, prose and judgment, and Fable "runs outside a panel only when the user named it for this task or two Opus 5.5 xhigh runs of the same brief failed its check" (poteto-mode/SKILL.mdSubagents). - orbit. The guide's "Run the lead on Fable" (
docs/guide/08-make-it-fit.md) runs the session on Fable athighormediumand addscode,judgmentandverifierrows onopuswith empty Effort cells. Spawns then run on Opus, and the lead still picks their level. The advisor stays on Fable. A variant putsjudgmentonfableathighto keep a second model in the panels. - Why. Orbit's own recipe, from the user's ask on 2026-11-11 to use Fable where it pays off. Anthropic's Fable 5.2 migration guide says early testers had good results with Fable "as the planner and orchestrator for other models", and that "subagents can be a large part of a session's cost". Fable 5.2 costs $8 and $40 per million input and output tokens, and Opus 5.5 costs $4 and $20 (Anthropic's model table, cached 2026-11-03). Empty Effort cells keep the lead's per-unit pick, because a row's effort is exact.
- pstack. "For panel roles (arena runners, interrogate reviewers) the value is a list, and one subagent runs per entry, alias entries included, so the list length sets the count." The arena judge comes from the
arena cross-judge poollist (skills/setup-pstack/SKILL.md). "Pick the runners from thearena runnersline ..., by defaultclaude-fable-5-2-high,claude-opus-5-5-high-fast, andgemini-4-neon-xhigh." (skills/arena/SKILL.md) - orbit. One list cell, the
other-familyModel. Interrogate runs arena's default seats as its two adversaries, one onjudgmentand one oncode. Arena, and the arena architect runs for a large or contested design, default to two runners, one on each role. Arena still spawns more runners for several design directions, and every runner runs onjudgmentorcode. Under T3 Code with a roster, each panel gets one seat on another family, theother-familyrole (Seats on another family). Its list orders the providers to try for that one seat, as pstack'sarena cross-judge poolorders the judge's, so its length never sets a count. The seat tries at most three of them (Ranked providers). Without a roster, the judge runs on the first of the two models that differs from the parent's, and arena skips it when neither differs. - Why. Every other Models cell holds one model. The
other-familylist is a preference order, so a seat whose first provider fails or shares the lead's family moves to the user's next pick, not to a provider the user never chose. When every pick fails, the seat is skipped, as pstack's arena skips its judge, unless the list ends withinherit. The adapter maps each default list toopusand the session model, which are thejudgmentandcodedefaults, so two seats keep pstack's default panels. A project changes a panel's models through those two rows. pstack's defaultarena runnerslist holds three slugs since bf0c7cd, two Claude and one Gemini. Orbit keeps two runners onjudgmentandcode, because every runner already runs on a Models role and a third default runner on Fable or a-fastslug has no role to map to (P13 plan gate, 2026-11-17). The one roster seat stands where pstack's non-Claude slug stands.
- pstack. Each seat below holds a non-Claude slug in its
setup-pstackdefault list.arena cross-judge pool: claude-opus-5-5-high-fast, gemini-4-neon-xhighandinterrogate reviewers: claude-fable-5-2-high, claude-opus-5-5-high-fast, gemini-4-neon-xhigh. The trail reviewer runs "on another vendor's model than the one that did the work" (skills/show-me-your-work/SKILL.md). At e1cc031howspawns one explainer, andwhyandreflectspawn nothing. de8ee15 put a grok slug on these seats and onhow explorer,reflect tooling,why investigatorsandswarm workers. - orbit. Under T3 Code with a roster, four seats move to the roster (T3 Code host). They are arena's judge, interrogate's second adversary, the show-me-your-work trail reviewer and autopilot-full's audit seat, one per full swarm. Each panel fills one roster seat, and only the agent whose turn end returns to the user fills it. A subagent or a child, such as an autopilot owner, takes the absent column.
delegate_taskgetsoptions: []when the model lists no matching effort, and the lead settles the child by itschildThreadId. Each seat brief says "You are a subagent of this thread." Eachdelegate_taskcall passes aclientRequestIdthat differs per seat, provider and round and stays the same on a retry of that call.t3_thread_sendonly answers or steers a running child, and every re-check or new round is a new call. When every listed candidate fails, the skill reports the skip line where the seat's result would go, and a skipped trail reviewer leaves the Attention section's first line asnot reviewed: seat on another family skipped. When the result holdsnext, the lead neither waits nor polls, works on, and ends the turn when no work is left. The seat's wake comes inside a running turn under a Claude lead, else as a new run, as Own-checkout children says. A child question reported byhasPendingQuestiongets an answer throught3_pending_requeston thechildThreadId, or goes to the user. The lead reads the result and settles the child only atworkStateresult_available. The reply names another thread by thelinkits result carries. When every verifier candidate fails, the lead sends the per-entry<entry> unavailable: <reason>lines withsend_to_userwhen it is present, before the question. - Why. A Grok child loads orbit from Claude Code's copy and may load the
rulesskill, which stops a subagent, so the brief names the child a subagent and it never acts as a lead (advisor, 2026-11-16). These are the seats where pstack runs another vendor and that only read or review. One seat per panel stands where pstack's one non-Claude slug per default list stands, and it caps the non-Claude children a run spends. A subagent's turn end returns to its parent, and adelegate_taskchild's wake lands on the lead's thread, not the subagent's, so a subagent never sees its seat finish (advisor, 2026-11-16).delegate_taskmerges a givenoptionsvalue over the inherited ones and clears them on[], so omittingoptionswould run the lead's options, not the provider default. T3's orchestrator MCP doc (c2eb8f3, read 2026-11-16, re-checked 2026-11-17) sets the rest. Each review round is a newdelegate_taskcall with a distinctclientRequestId, stable across retries, and a call without one creates new work on retry. "When the result includesnext, completion wakes the parent: do not wait on the task or poll it." A child question goes to the parent, which answers witht3_pending_requestor asks the user, because "Users rarely watch child threads". Alinknever goes stale when the user renames a thread, as an id or title would.task_statussays "A finished turn with live child work is not done" (tool description, 2026-11-16). A fallback to another provider is a different call, so it takes a newclientRequestId, or the retry key would return the failed task (inferred). T3's system prompt rules out a re-check throught3_thread_send(Ranked providers). A Claude lead in the live matrix wrote one summary sentence before itsAskUserQuestionin place of the per-entry lines. T3's runtime prompt says text before a tool call may reach the user only as a one-line summary, and to usesend_to_userfor content the user must read in full mid-turn (#72).
- pstack. "spawn a subagent on another vendor's model than the one that did the work" (
skills/show-me-your-work/SKILL.md). de8ee15 read "a different model family from". - orbit. Keeps "a different model family from". Under T3 Code with a roster, the reviewer is the roster seat (Seats on another family). Otherwise it is an
Agentcall withmodelset to a different Claude tier, Sonnet under Opus, else Opus. The Attention header readsreviewed by <model>, as pstack's does, with the provider when the roster seat ran. When every listed entry fails, the seat is skipped and the section opens withnot reviewed: seat on another family skipped, and the lead never reviews its own trail in the reviewer's place. - Why. The roster seat is another vendor, as pstack asks. Without a roster only Claude models run, so "another vendor" cannot hold, and the different Claude tier is the closest reviewer. Inheriting the session model would make the reviewer the same model as the worker, so this spawn keeps an explicit tier. Orbit's header once added
(same family: no other available), and the user kept running into that caveat ("keep hitting Attention. Reviewed by Sonnet", transcript audit, 2026-11-10), so the header matches pstack again.
- pstack.
interrogateruns "Phase A steps 3 and 4 and Phase B of the arena skill", reads "theinterrogate reviewersline in place ofarena runners", and spawns "each adversary asgeneralPurpose". "A finding survives only when a run proves it." It never applies fixes (skills/interrogate/SKILL.md). - orbit.
/orbit:interrogatereads arena'sSKILL.mdby path and runs those steps with arena's default seats. Its two adversaries aregeneral-purposeagents, one withmodel: "opus"(thejudgmentmodel) and one on thecodemodel, each witheffort: "high". Under T3 Code with a roster, the second adversary is the roster seat instead, spawned withdelegate_taskisolation: "worktree", at the effort the adapter's effort rule maps. Its brief holds the commit to check out and the patch file's path, and it reports its worktree's absolute path and starting branch with each finding. Verify reruns each command in that worktree, and cleanup settles its thread and removes the worktree and that branch. Under a T3 lead withdelegate_task, an adversary that is not the roster seat runs as a same-providerdelegate_taskchild in the same way, because it keeps its repros in its own checkout (Own-checkout children). Inside a child, or withoutdelegate_task, the adversaries run one at a time in the agent's own checkout. Feature step 7 runs it when the design is contested. Known issues get fixed before the adversaries spawn (Fix before audit). - Why. Orbit has no
interrogate reviewersline, and the adapter maps a non-Claude slug to the session model, thecodedefault, so arena's default seats keep pstack's default panel. pstack's defaultinterrogate reviewerslist has a Gemini entry, so with a roster that seat goes to another family. It is the one roster seat that runs code, so it gets a worktree, whichdelegate_taskstarts at HEAD of the thread's bound checkout (Own-checkout children) and the seat reports back (T3-3 lane 1, 2026-11-16). T3 starts that worktree on a new branch, and a seat that checks out a commit detaches from it, so the lead deleted nothing and the branch stayed (T3-3 lane 1, 2026-11-16). Arena's step 3 namesorbit:poteto-agent, so interrogate overrides it withgeneral-purpose, the adapter's mapping forgeneralPurpose. An adversary reasons about how code fails, which is judgment work, and rules.md picks high for that.
- pstack. The judge comes from the
arena cross-judge poolline, "by defaultclaude-opus-5-5-high-fast,gemini-4-neon-xhigh. The judge is never from the parent's vendor, and Fable counts as Claude for that check". "If the Task tool rejects a slug, use the nearest valid slug of the same family, then the next such entry, or skip the judge, and say so." (skills/arena/SKILL.md) - orbit. Under T3 Code with a roster, the judge is the seat on another family in any session, briefed with the rubric and each candidate's absolute path (Seats on another family). Without one, the judge's model comes from the
judgmentandcodemodels and is never the parent's model, so a different model stands in for a different family. It runs on the first of them that differs. When both resolve to the parent's model, as in an Opus session under the defaults, arena skips the judge, Claude picks the base alone, and the synthesis note says so. If theAgenttool rejects the model, arena also skips the judge and says so. With the parent's model unknown, it uses thejudgmentmodel and says the model check did not run. A session-family judge is a foregroundAgent, so it returns before Phase D's read starts. Only the roster seat, an asyncdelegate_task, scores while the lead reads. - Why. With a roster, pstack's "The judge is never from the parent's family" holds again. Without one only Claude models run, so a literal copy of that line would skip the judge in every session. A different Claude model is the closest stand-in and the fallback, as the Trail reviewer takes a different tier. A judge on the parent's own model would share its blind spots, so arena skips it rather than run one (inferred). Trying
judgmentfirst is orbit's choice. pstack's default pool names a-fastslug and Gemini, which map to no Models role, so the judge stays on the roster seat or the first ofjudgmentandcodethat differs. Arena Phase C's "While you read in Phase D, the judge scores each candidate" assumed a background judge, and T3 allows "Background agent:delegate_task, never a backgroundAgent." (orchestrationInstructions.ts:9in pingdotgg/t3code at c2eb8f3), so only the roster seat reads in parallel with the lead (T3 clash audit, 2026-11-17).
- pstack. "Copy the scope's uncommitted changes into each worktree." Verify reruns each finding's command. The scope runs "through the working tree" (
skills/interrogate/SKILL.md). - orbit. The scope includes untracked files (
git status --short). The lead names the commit each adversary checks out and saves the scope's uncommitted changes, untracked files included, as a patch in a freshmktempfile that the adversary applies first. Verify reruns each command in that adversary's own worktree, whose path theAgentresult or the child's report returns. Then the lead removes every adversary worktree withgit worktree remove --force <path>, deletes its branch and deletes the patch file. - Why. Claude Code creates an
isolation: "worktree"worktree only when the agent spawns, so the lead can't copy files into it beforehand. A worktree with changes stays after its agent returns, at its own path, so the repro files are there for the rerun and nothing removes them unless the lead does. Removing a worktree leaves itsworktree-agent-*branch, and a fixed patch path would collide between sessions (both seen in a live run). A plaingit diffleaves out untracked files, which are part of the working tree (inferred).
- pstack. No rule. An adversary gets "the scope, the intent, and this brief" (
skills/interrogate/SKILL.md), and project context comes from the harness. - orbit. rules.md's Review line passes the Workflow block's "Review also checks" sources and the
Before codeentries for the changed paths to/orbit:interrogate. Its Fan out puts them in each adversary's brief. - Why. The rule started when interrogate's reviewers were
Exploreagents, which don't load CLAUDE.md. Adversaries aregeneral-purposenow, so that reason no longer holds. The sources stay in the brief for the reason in Project sources in briefs. A path-scoped source reaches the agent by name, not through AGENTS.md pointers.
- pstack. No such step. Interrogate says "Report the findings. Do not apply fixes." (
skills/interrogate/SKILL.md). - orbit.
/orbit:interrogatehas a Fix first step between Frame and Fan out. It fixes open PR review comments that hold up against the code, confirmed bugs and Check failures. A freshorbit:poteto-agentmakes each fix under the Delegate gate and gitops commits it. Then the Check runs, and the adversaries check out the commit after the fixes. A suspicion the user wants tested stays for the audit. The step skips code the user plans to replace, a design or plan with no code, a branch this session does not write, and a scope with uncommitted changes. Interrogate still never fixes its own findings, and its Output reports the fixes or why the step was skipped. - Why. tstack's council skill (t3dotgg/tstack
skills/council/SKILL.md) says "Before an audit, address outstanding PR comments, fix known issues, and run relevant checks. Give council the updated implementation; do not defer known fixes until reviews return." Without the step, adversaries spend a run re-proving known bugs (inferred). The user asked for it on 2026-11-16, with the rule that a reviewer never writes code and the code agent never reviews, so poteto makes the fixes. Babysit already routes fixes through the Delegate gate (playbooks/babysit.md). Only comments that hold up get fixed, because rules.md asks for skeptical triage. Suspicions stay for the audit. The step skips a dirty scope because gitops commits by path and would sweep the user's uncommitted work into a fix commit (advisor plan gate, 2026-11-16). It skips a branch this session does not write because an autopilot owner is its branch's only writer.
- pstack. No rule. Delegates rely on the harness loading AGENTS.md.
- orbit.
skills/setup/workflow-block.mddefines aBefore codeline of path-scoped entries, each a file or a skill-load command. The Delegate gate copies the matching entries verbatim into each code delegate's brief.agents/poteto-agent.mdreads each file whole, runs each command and follows any "read X first" pointer. - Why. Orbit's choice. Path-scoped guidance, such as a library's agent guide, reaches each delegate by name, so no delegate depends on finding it through AGENTS.md pointers.
- pstack. No handoff line. Feature says "Stack follow-ups" (
playbooks/feature.mdstep 6), and Autonomy proceeds on reversible work (poteto-mode/SKILL.md). - orbit. Autonomy's four paragraphs sit in rules.md Task End, after "finish every follow-up inside the goal yourself" and "List anything outside the goal. Don't build it." It ends with one
Next:status line. A change that waits on the user's review gets a line that names the review and gives/orbit:shipas its own step after it. Landing a held change never folds into an/orbit:startfollow-up. - Why. Orbit has no Autonomy section, so its lines sit with the other task-end rules. The user's global CLAUDE.md asks for the
Next:line. The goal rules are orbit's choice. They finish work inside the goal and list scope creep instead of building it. A follow-up line that bundled "push and merge" started a merge the user had held for review (transcript audit, 2026-11-10), so landing stays a separate step the user runs.
- pstack. No such section.
- orbit. rules-shared.md Code Quality items 1 to 6, kept from orbit's team config (6af3e88). Item 5 keeps only its first sentence, "Fix the cause, not the symptom."
- Why. User-authored config that predates the port, kept as written. Item 5's second sentence, "No workarounds without asking", conflicted with never-block-on-the-human, so review cut it (49e6804).
- pstack. No such section.
- orbit. rules-shared.md Files keeps
.mdfiles out of a repo root, puts working notes in auto memory, and puts permanent docs indocs/. - Why. User-authored in 6af3e88. Notes moved from
.claude/to auto memory because Claude Code prompts on every.claude/write.
- pstack. No such default. CI on the PR gates landing.
- orbit. rules.md Workflow Defaults sets Check to
vpr check, thenvpr buildandvpr testif those scripts exist. The flows' verify steps run it, and under LandmainShipping runs it in place of CI. - Why. User-authored in 6af3e88. Under Land
mainno PR opens, so no CI runs.
- pstack. Interrogate's Frame says "State the intent with the in-your-own-words skill". That skill is one sentence, "Restate in your own words what you think the user's goals are and what problem they're trying to solve." (
skills/in-your-own-words/SKILL.md) poteto-mode's Non-negotiables also open with "First message → restate the user's goal and the smallest change that meets it, then proceed." (poteto-mode/SKILL.md). de8ee15 opened them with "Your first message on a task → tell the user, in your own words, their goal", added in upstream commit ab1f702. - orbit. Interrogate's Frame inlines the sentence, as "State the intent by restating in your own words what you think the user's goals are and what problem they're trying to solve". rules.md's Remaining triggers open with the de8ee15 rule, copied as pstack worded it (Trigger list).
- Why. Orbit doesn't ship in-your-own-words, and one sentence carries it, as ketchup inlines
bro's tone. poteto-mode's rule says "in your own words" in plain text and names no skill, so it needs no inlining. It was left for an upstream bump (docs/trails/pstack-parity.tsv), and the de8ee15 pin brought it in.
- pstack. Every change lands through a PR. Opening a PR is "Invoked at the end of every other playbook" (
playbooks/opening-a-pr.md). - orbit. With no Land line, Land is
main(opening-a-pr.mdLand,skills/setup/workflow-block.md). Opening a PR then pushes nothing and opens no PR. Babysit says there is no PR and stops. Shipping fast-forwardsmainbottom-up after the verifier and the Check line pass. A stack is a chain of unpushed task branches, and measurements go in the reply. rules.md's "per-PR verdict" then means one verdict per task branch. - Why. The user's product call. A solo repo needs no PR. Land
prkeeps pstack's PR path for projects that want it.
- pstack. A task ends at Opening a PR. "Opening a PR does not start a babysit" (
playbooks/opening-a-pr.md), and "Babysitting never authorizes merging" (playbooks/babysit.md, last bullet). - orbit. A change task ships once the verifier passes it. It runs Opening a PR, then Babysit
driveunder Landpr, then Shipping (opening-a-pr.mdShip by default). "don't ship", "I'll land it" or "stack them, don't ship" stops it at the PR or branch. Ship by default grants a multi-phase plan landing authority, so the plan runsautopilot-full(Multi-phase plan). The opt-out withholds that authority, so the plan's go runsautopilot-stack. rules.md's Babysit trigger keeps "Never triggered by merely opening a PR". - Why. User preference, 2026-10-29: "ill be writing more ship it when verified. than if it was dont ship until i say so". Where the Babysit trigger and Ship by default meet, the recorded preference wins. The standing request starts Babysit, not the PR.
- pstack. "for irreversible writes: force-push to shared branches, deploys, data deletion, customer messages" (
poteto-mode/SKILL.mdAutonomy). - orbit. The same list in rules.md Task End. A product call goes to the advisor first, and the lead pauses only when the advisor's pick drops or changes what the user asked for (Test before asking). It adds "Pushing, merging and opening PRs are not irreversible."
- Why. A change to the goal is the one call that neither an experiment nor the advisor can make for the user. The old line paused for every product call, and the lead asked the user calls the advisor could make (Test before asking). Ship by default pushes and merges, so the rule says those don't pause.
- pstack. e1cc031 has no such trigger. de8ee15 said "fix it in its own PR. Don't block." (
poteto-mode/SKILL.md) - orbit. In orbit's own repo, a broken orbit skill is fixed in its own PR. Under Land
main, that is its own worktree and branch, with its own Opening a PR run (rules.md Non-negotiables). From any other repo, the lead files an issue on orbit's repo (gh issue create --repo aayushbtw/orbit) with what broke, the session and a verbatim quote, then keeps going. It doesn't edit orbit from there. - Why. Under Land
mainno PR opens, so a separate branch keeps the fix out of the task's change. Orbit can fix only the skills in its own repo (inferred). From another project, an edit lands in the installed plugin cache or pulls the session out of its task (inferred), and the user said "dont edit orbit just open issue" (transcript audit, 2026-11-10).
- pstack. e1cc031 cut the subagent paragraph from
playbooks/opening-a-pr.md. Autopilot-full step 2 still says "The code is ready only after the slop-strip ... and the/no-commentspass". - orbit. Opening a PR's Ship by default says a subagent that opens a PR returns without babysitting, and an autopilot owner babysits and stops where its brief says. The owner's
code-readytrail row lives in autopilot-full step 2. The same paragraph says no child watches a PR, and a child that did PR work ends with the PR line, per babysit's Host watcher. In T3 Code, an owner's build round reads its PR once withwatch-pr --onceand returns with its PR line, and the root babysits (Own-checkout children). There a child, meaning any subagent, autopilot owner, gitops agent ordelegate_taskchild, never watches, even when its brief says babysit. In T3 it runs one--onceread per push, acts onBLOCKERand returns on any other verdict. Its last line isPR <full URL>, head <SHA>, checks and reviews <state>. - Why. A subagent reports once, so code-ready is a trail row (Autopilot-full). T3 hides
watch_pull_requestandunwatch_pull_requestfrom subagents (McpInvocationContext.hiddenFromSubagentsin T3'spr-tools.tsat c2eb8f3). T3's source-control doc says "Subagents cannot watch pull requests; the thread that delegated to them does." T3's runtime rule (runtimeInstructions.tsline 3) says "Delegated children never watch, even if the brief says babysit. After PR work, a child's last line is: PR , head , checks and reviews ." So the child rule wins over a brief that says babysit to green.
- pstack. "One verifier per PR, each a Cursor cloud agent, runs the change the way a user would" with the matching control skill, parent against head, and posts its verdict on the PR (
playbooks/shipping.mdstep 1). - orbit.
agents/verifier.md, a freshorbit:verifierthat shipping step 1 (/orbit:ship) or the autopilot-full root spawns withisolation: "worktree". Shipping spawns it with theverifierrow's model and effort, not thecodemodel (Model per role). It checks out the head SHA detached, runs parent against head, and returns its verdict. gitops posts it withgh pr comment, with the SHA and the patch-id. It has Write but no Edit tool, so scratch scripts go to/tmpinstead of a Bash heredoc, which the worktree guard refuses (rules-shared.md "One plain command per Bash call"). It drives prompt changes withclaude -p --plugin-dirin a scratch repo. Its prompt carries the Machine line. Its tools addToolSearchand two named T3 read tools,list_thread_pull_requestsandt3_thread_read. It has nopreview_*tool, and its body says never to verify through T3's preview browser, so it drives a UI through the control skill's own headless browser. Shipping step 7 removes its worktree once the branch lands. Under a T3 lead withdelegate_task, the native verifier is a same-providerdelegate_taskchild withisolation: "worktree"and the same brief, plus the agent file and itstoolslimit (Own-checkout children). Before any verifier spawns, the advisor reviews the diff (Diff review). Under T3 Code with a roster and averifierrow ofother-familyorother-family, inherit, the lead spawns each PR's verifier as a seat on another family throughdelegate_taskwithisolation: "worktree", one child thread per PR, under the Parallel cap. Autopilot owners and swarm lanes stay on the lead's provider. The brief namesagents/verifier.mdto read first, states itstoolslimit, carries the SHAs, intent, changed paths and skill path, and asks for the worktree's path and starting branch so the lead can remove them once it settles the child. A child that cannot run the change repliesBLOCKED: <reason>, and the lead names that entry unavailable in the reply and tries the next candidate. When none is left, the native verifier runs only if theverifierModel isother-family, inherit. Otherwise the lead writes one<entry> unavailable: <reason>line per entry before any question or hold. Then it asks the user before a native verifier runs, with the native verifier offered but not marked recommended, and with no user to ask it holds the PR. AFAILstands. The lead settles each child thread after reading its verdict. - Why. The adapter maps a cloud agent to
isolation: "worktree". A subagent returns its verdict as text, and gitops runs every forge write. A prompt change has no app to run, so a headless session drives it. A cloud agent's VM goes away on its own, and a local worktree stays until something removes it. Before it had T3 tools, the verifier saw none in a T3 session, because itstoolslist allowed only Read, Grep, Glob, Bash and Write (T3-native plan, Appendix A). With T3 tools allowed, the verifier loaded and calledt3_preview_listin the T3 probe lane on 2026-11-16. T3 then ruled preview out for this job. Its system prompt says "Never usepreview_*unasked to verify or show your own edit." (orchestrationInstructions.ts:24in pingdotgg/t3code at c2eb8f3) and "When browser work was asked,preview_*is the browser you share with the user" (line 36). A verifier is never asked for browser work by the user, so the preview tools left its list. The rule also sits in its body, because a seat on another family reads the file by hand and a hand-readtoolslist binds nothing (Seats on another family). A verifier only judges, so it gets no T3 write. The list once held the wildcardmcp__t3-code__*. Under a stubt3-codeserver, a verifier with it loadedt3_project_delete,delegate_task,t3_thread_launch,schedule_taskandunlink_pull_requestthroughToolSearch, and gitops's named list blocked the same probe (interrogate rerun, 2026-11-16). So the verifier's list names each tool too. The seat follows the operator's directive and the advisor plan gate of 2026-11-16 (Model per role). pstack's verifier runs on theswarm workersmodel. At de8ee15 that was a grok slug by default, so another family checked a Claude author's change. e1cc031 moved the default toclaude-opus-5-5-high-fast. A child thread loads no orbit agent, so the brief names the file and states its limit, as Where a child runs step 5 says (Own-checkout children). Another provider can lack the project's toolchain orclaude -p, and aFAILthere would hold a correct change, soBLOCKEDhands the run on instead (inferred). T3 forbids a fallback to native tools unasked, and without a verifier nothing lands, so the lead asks, and a trailinginheritrecords the user's yes (Ranked providers). Only theverifierrow'sinheritcounts. Setup writes it only after a yes to its second question, and a user can end theother-familylist withinheritand still answer no (advisor, 2026-11-17). In the liveverifier-all-faillane at 31ee0b3, the lead merged both entries' reasons into one question sentence, wrote no per-entry line, and marked the native verifier recommended, which theverifierrow had withheld. So the adapter and shipping step 1 now spell out both rules. A subagent never sees its child's wake, and a child starts no child thread (orchestrationInstructions.ts:13), so owners and lanes take no seat on another family (Seats on another family).
- pstack. No diff review. Shipping step 1 says "One verifier per PR, each a Cursor cloud agent, runs the change the way a user would" (
playbooks/shipping.md). "Verification always runs on theswarm workersmodel,claude-opus-5-5-high-fastby default." (poteto-mode/SKILL.mdSubagents) - orbit. Shipping step 1 briefs
orbit:advisorfor a diff review of each PR before any verifier spawns, on Fable from its frontmatter or theadvisorrow. The brief holdsgit diff <parent>..<head>, the intent paragraph and the Plan gate's project sources for the changed paths. It asks for blocking findings, meaning bugs, broken project rules and missed cases. The lead checks each blocking finding against the code and dismisses one that doesn't hold, with its reason. Each one that holds gets a fix round through the Delegate gate. Under Landpr, gitops pushes the fix, and Babysit brings the PR and every PR stacked above it back to merge-ready before any verifier spawns. The review covers the patch it read. It runs again, ongit diff <reviewed-sha>..<head>, only when step 3 finds that patch changed, or after a fix for aFAIL. Step 3's "fresh checks and a review" becomes "fresh checks and step 1's diff review". Non-blocking findings and the model that reviewed the diff go in the reply. The review is not a verdict and never replaces the verifier. Under T3 Code the advisor is a foregroundAgent. Only the agent whose turn end returns to the user runs the review, so an autopilot owner or swarm lane never does. Autopilot has no diff review, because its merge round runs only shipping steps 3, 4 and 7 (skills/playbook/playbooks/autopilot-full.mdstep 5). This is a known gap.agents/verifier.mdstill names no model. - Why. The verifier runs on the session model, so in an Opus session it is the same model as the code's Opus author. The user asked for a Fable code review by hand 6 times (transcript audit, 2026-11-20). Fable's read-only diff reviews then found real bugs that the verifier had passed, traced from the code, not from a run. A
model: fableline inagents/verifier.mdwould not close the gap under T3, because there the verifier runs as a child thread that loads no agent frontmatter (Own-checkout children). The advisor already runs as a foregroundAgenton every host, on Fable. The advisor reads code and never runs it, so the lead checks a finding before it spends a fix round on it. A verdict on an unpushed fix would not describe the head that lands, so the fix is pushed and the stack made merge-ready first (inferred). A fix round before the verifier keeps a verifier off a head the review already blocked (inferred). A second review of an unchanged patch would read the same code again, so it reruns only on what changed.
- pstack. Step 1 gives the verifier no written brief.
- orbit. Shipping step 1 writes the change's intent in one paragraph from the goal and the code. The verifier's brief holds it, the head and parent SHAs, the changed paths and the verify or control skill path.
- Why. Orbit-only. A fresh subagent sees only its brief, so the intent tells it what a user should see.
- pstack. Merge "with
origin pr merge <pr>orgh pr merge <pr> --squash", then read backstate,mergedAt,baseRefName. A child moves onto the merge result and pushes "with--force-with-lease" (playbooks/shipping.mdstep 4). - orbit. gitops merges with
--delete-branchand the Land line's method,--squashby default, and reads back the same fields. It moves the child withgit rebase --onto origin/<trunk> <tip>, where<tip>is the merged PR'sheadRefOid. - Why. The merge method is the project's choice in its Workflow block (
skills/setup/workflow-block.md). After a squash, a plain rebase replays the parent's commits and conflicts.
- pstack. "Force-push only your own branch, with
--force-with-leasepinned to the SHA you last pushed or took over" (playbooks/autopilot-full.mdstep 2). The owner rebase in that step and shipping step 4 push with a bare--force-with-lease. - orbit. gitops makes every force-push as
git push --force-with-lease=<branch>:<sha>, with the SHA the lead passes. It rewrites pushed history only on a task branch the session pushed or took over, and stops if the remote tip differs (agents/gitops.md). - Why. A bare
--force-with-leasechecks the remote-tracking ref, which any fetch moves, so it can overwrite a push nobody saw. gitops starts fresh each time, so only the lead knows the SHA.
- pstack.
playbooks/opening-a-pr.md, three paragraphs. "Work from a git worktree off trunk, and keep unrelated work out of it. Rebase into small, landable commits. Run/deslopfromcursor-team-kitover the diff before commit." PRs open "through the run's PR tool withdraft: false. Without one, usegh pr createwithout--draft, ororigin pr create --status openwhen Origin's CLI resolves the repository." - orbit.
skills/playbook/playbooks/opening-a-pr.md, the last step of every change flow (Opening a PR gate). A Land paragraph decides whether a PR opens (Land default). gitops commits each reviewed unit by path and runs the rebase, with fixup and autosquash (Commits). It keeps "/orbit:no-commentsbefore review", which upstream cut. The Body sentence adds that evidence logs, run histories and long tables go in a file or a PR comment that the body links to.ghis the only forge. gitops opens the PR withgh pr createwithout--draft, then links it to the session's thread per the host adapter, and skips the link wherelink_pull_requestis absent. Under Landmaina stack is a chain of unpushed task branches. A Ship by default paragraph and a Reply: line follow. - Why. Origin and Graphite are not vendored. pstack's adapter maps a built-in PR tool to the forge command. Orbit's T3 adapter maps it to
gh pr create, thenlink_pull_requestwith the PR URL (T3 Code host). Orbit'sno-commentsskill says Opening a PR runs it before review, and autopilot-full step 2, orbit's and upstream's, gates the code-ready head on the/no-commentspass, so the clause stays. e1cc031 moved the rule that evidence goes in a linked artifact, not the body, fromopening-a-pr.mdintoworker.md, which orbit does not ship (Trigger list), so orbit keeps the rule in Opening a PR.
- pstack.
playbooks/babysit.md, six bullets. "Useghby default, ororigin prwhen Origin's CLI resolves the repository. Never require Graphite (gt)." Modes aredrive,threads-onlyandcheck. "Never rebase, retarget, or force-push." It says nothing about a host's own PR watcher, and the Cursor-only features table in pstack's Claude Code adapter (pstack/docs/harness/claude-code.md) lists only Cursor Automations, Bugbot, cloud agents and the built-in PR tool. Its watcher reads empty checks as a retryablechecks-unavailableerror (scripts/watch-pr/github.ts) and, after five query errors, stops atBLOCKER: status-query, so it never reportsREADYon a repository with no CI. Both still hold at e1cc031. - orbit.
skills/playbook/playbooks/babysit.md. The stance line adds that a request to land or ship is Shipping (/orbit:ship). A paragraph after it says the user's request or Ship by default starts it, indrive, even for a small PR. Under Landmainit says there is no PR and stops. Fixes go through the Delegate gate. The forge bullet keeps "Useghby default." and drops the Origin and Graphite clauses. Bugbot replies go through gitops with thegh apicommand, with the payload written by the Write tool. The last bullet adds orbit's standing ship default. The watcher is pstack'swatch-prinskills/playbook/scripts/, with three fixes. First, when both check sources are empty, it asks GitHub whether the repository runs checks. With no workflow on the PR head or the base branch that runs on apull_request*event or on a push that can reach a PR branch, no check on an earlier commit of the PR, no check on one of the repository's 10 most recently updated pull requests, and no protection rule or ruleset that requires checks, the read isno-ciand the PR can reachREADY. Second, under--status-onlya query error returns thestatus-queryblocker at once, with no retry. Third,--oncemakes one classifying read and never sleeps.READYandBLOCKERcome as usual. A pending PR ends atTIMEOUTwith reasonpending-checks, which names itsfrontier, orqueued-stackwith--queued-stack, exit 5. A retryable query error ends atTIMEOUTwith reasonstatus-unavailable, even under--max-query-errors 1, and a non-retryable one at thestatus-queryblocker. That timeout'sprnames the PR whose read failed. Agherror that says the PR or repository does not exist, or that the branch has no PR, is the non-retryablenot-found. Thequeued-stacktimeout says inwaitingOnwhether its frontier waits onpending-checks,review-requiredor only themerge. The watcher readsreviewRequests, so a READY gate and areview-requiredwait name the requested reviewers.--onceconflicts with--status-only. A**Host watcher.**bullet sets the lead's wait by host. The session is in T3 when anymcp__t3-code__tool is in the tool list. In T3 withwatch_pull_requestpresent, the lead pushes any branch update first, and reads once with--once. When the verdict waits on checks, it callswatch_pull_requeston the PR the verdict names and ends the turn. That is apending-checkstimeout'sfrontier, astatus-unavailabletimeout'sprwhen itsfailure.kindischecks-unavailable, or aqueued-stacktimeout'sfrontierwhen itswaitingOnispending-checksorreview-required. Any otherstatus-unavailablefailure gets no watch, so the lead reports it, hands back and reads again with--onceon its next turn. When a merge was asked for, aREADYwithreviewDecisionREVIEW_REQUIREDor achanges-requestedblocker waits on a review, so the lead names the reviewers, watches and ends the turn. A wait only on a merge someone else holds, such as a merge queue, auto-merge or a person, gets a watch, and the lead says what the PR waits on and ends the turn. Every watch passesuntil. It is"merged"when a machine holds the merge, such as a merge queue or auto-merge, or when a merge was asked for and the PR waits on checks or a named reviewer. It is"checks"on every other wait, such as a merge a person holds or a review with no reviewer requested. In Shipping step 5 under T3, a frontier that waits on the lead's own next step-4 merge is no handback, so the lead runs that merge and reads again. On a wake it reads again with--once, watches a new frontier the same way, and callswatch_pull_requestagain only after T3 says it stopped watching a PR that is still open and still waited on. A tool error takes the tool-off path. In T3 with the tool off, the lead reads once per turn and hands back, with no sleep, loop,gh --watch,MonitororScheduleWakeup. Outside T3,driveruns the bare watcher under/loopwith no interval, per the adapter's/looprow. A handback is aREADYthat waits on nothing, a conflict report, a question to the user, the ceiling or a stop. Before one, the lead callsunwatch_pull_requeston each PR it watches. The bullet's second paragraph says no child watches a PR (PR owner checks). Shipping step 5 waits per Host watcher, and step 6 unwatches at the ceiling. A Reply: line names the mode, the verdict, each thread, what a wait waits on and each PR unwatched. rules.md's Babysit trigger drops the clause about Cursor's built-in babysit skill. The mode bullet keeps de8ee15's request-to-mode mapping, which e1cc031 cut. "Check on X", "is it green" and "anything outstanding" ask forcheck, "babysit this", "get it green" and "wait on CI" ask fordrive, "address the review comments" asks forthreads-only, and an undeclared mode isdrive. rules.md's Babysit trigger says the playbook's first bullet owns that mapping. The forge bullet names the watcher's command form,watch-pr --owner <owner> --repo <repo> --pr <n>, with--stackfor a stack and--status-onlyfor one read. - Why. Claude Code has no built-in babysit skill, and Origin is not vendored. The bold
**Host watcher.**lead-in keeps a citable anchor, because upstream's bullets have no step numbers and shipping step 5 and the T3 adapter cite the rule. gitops runs every forge write and can write a file only through the shell. Shipping needs the watcher's terminal verdict, so a small PR still runsdrive. pstack's watcher never reaches a verdict on a repository with no CI and stops atBLOCKER: status-query, and orbit's own repository runs no CI. A push workflow attaches checks to a same-repository PR branch, so it counts unless itsbranchesfilter names only the base branch or it runs only on tags. A workflow that runs only on push to the base branch, such as a deploy, never attaches a check to a PR, and it sets the base branch head's rollup, so neither its file nor that rollup counts. A status-only read promises one table and an exit, but pstack's ran the full retry schedule on a query error, about 12 minutes on one PR. A host PR watcher such as T3 Code'swatch_pull_requestwakes only on events. On a PR with no CI and no reviewer, an agent that armed it in place of the watcher waited on nothing. T3 now wakes a GitHub watch once 15 minutes pass with no check (PULL_REQUEST_WATCH_NO_CHECKS_MSinapps/server/src/orchestration-v2/pullRequestWatch.tsat 6d05c5f), and never off GitHub. So the--onceread comes before the watch. T3's runtime rule for every thread (runtimeInstructions.tsline 3 in pingdotgg/t3code at c2eb8f3) says "Call watch_pull_request when asked to monitor or babysit, or when work you were asked to do, such as a merge, still waits on checks or reviews, even after an earlier unwatch. Finish any needed branch update and push first, then watch and end your turn. Otherwise act now or hand back. Delegated children never watch, even if the brief says babysit. After PR work, a child's last line is: PR , head , checks and reviews . Links and watches persist across pushes. For PR waits, no sleep, loops, gh --watch, Monitor or ScheduleWakeup. Before handback, call unwatch_pull_request on any PR you watch." The old--timeout 120wait slept between polls, which that rule bans. The tool's own description in T3'spr-tools.tssays "After this call, end your turn. The watch stays on after each wake, so do not call this again on a wake. A wake is not a merge decision. Check readiness yourself."--status-onlycould not be the one read.runSimpleinpolicy.tsreturnsSTATUSwith reasonstatus-onlybefore it classifies the PR, so it never reportsREADYorBLOCKER.--timeout 0turns the deadline off, so the watcher polls forever. So--oncemakes the deadline pass on the first read. A read right after a push often finds no checks yet, a retryablechecks-unavailableerror. When--onceended every query error at thestatus-queryblocker, that read stopped babysit before it armed a watch (P6 audit, 2026-11-17), so a retryable error ends at astatus-unavailabletimeout, which arms one. Onlychecks-unavailableclears when a push's checks register, which wakes a watch. A failedghcall or an unparsable reply is no change to the PR, so a watch on it would never wake (P6 audit, 2026-11-17).--pr 999999and a branch with no PR came back as retryablecommand-exiterrors, so--onceended them atstatus-unavailableand babysit would watch a PR that does not exist (P6 verifier, 2026-11-17). They now end at thestatus-queryblocker, as the long-running watcher already did for a branch with no PR. In--stackmode the read that fails can be an upstack PR, so the timeout names it. T3's tool description lists no merge among its wakes. Orbit read that as no wake on a merge, so a lead whose PR waited on someone else's merge handed back. T3's source-control doc says "When the pull request merges or closes, watching ends, and the agent is woken once unless it merged or closed the pull request itself with the host's CLI. A reopened pull request needs a new watch. Watching also ends after 10 wakes in a row that bring only comments, after 8 failed reads in a row, or when you press Stop on the thread." (docs/user/source-control.md:345-348in pingdotgg/t3code at 6d05c5f).PullRequestWatchReactor.tsskips that wake only for a merge or close command in the thread's own turn, and says "Anything ending later, as a--automerge does, is news." The merge or close wake fires whateveruntilsays (finished()inapps/server/src/orchestration-v2/PullRequestWatchReactor.ts).untilonly decides whether the thread shows as Working (packages/shared/src/orchestrationV2PendingBackgroundWork.ts). The tool's text says "Omit it to keep the current value; a new watch starts at "checks". "checks": until checks finish, for a PR you hand to the user. "merged": until it merges or closes or you unwatch it, when you will act on it yourself, such as merging after someone else approves." (apps/server/src/mcp/toolkits/pullRequests/tools.ts). So a wait on someone else's merge keeps a watch and ends the turn. A merge queue or auto-merge holds the thread in Working with"merged". A merge a person holds passes"checks", so the thread returns to the user's inbox and the merge still wakes it. Omittinguntilwould keep an earlier"merged", so every call passes it. The lead re-arms only after T3 ends a watch on a PR that is still open. T3 also wakes when the branch needs a rebase, and when every check passes on a PR with no required checks (apps/server/src/orchestration-v2/pullRequestWatch.ts). The installed T3 Code (Nightly) v0.1.11-nightly.20261119 holds the end message "so T3 Code stopped watching it.", theuntiltext, the tool description, the rebase and all-checks wake messages, the 10-update and 8-read limits and the 15-minute no-checks constant (rg -aon itsapp.asar, 2026-11-20). A live probe on that build confirmed the merge wake on 2026-11-20. A Claude lead watched aayushbtw/orbit-t3-lane-ci#19 withuntil: "checks"after its only check had passed, then ended its turn. A detachedgh pr mergethat the lead had started before ending its turn merged the PR at 05:44:06Z. T3 woke the idle thread by 05:45:11Z with "Pull request #19 (https://github.com/aayushbtw/orbit-t3-lane-ci/pull/19) was merged into main by aayushbtw as ccfef9e, so T3 Code stopped watching it." A wait on a review arms a watch. T3 says "Otherwise act now or hand back", and in Shipping the merge the frontier waits on is the lead's own step 4, so it acts. In--stackmode the PR with pending checks can sit above the one named, so the timeout names itsfrontier. The host key is anymcp__t3-code__tool, the same key rules.md uses forrun_in_background: false, because T3 can turnwatch_pull_requestoff and leave the rest on. Outside T3 no such rule applies, so the bare watcher under/loopstays (inferred). rules.md routes "check on PR X" and "anything outstanding on X" to Babysit and asks for a mode before any poll, so the flow has to say which mode each request gets. Two headless leads passed the watcher a PR URL and got exit 64, its usage error (P13 audit, 2026-11-17), so the bullet names the flags. In a T3 lane asked to have gitops wait on CI, the lead tookcheck, gotTIMEOUT pending-checksfrom gitops and armed no watch (P6 lane B, 2026-11-17). T3 says "Call watch_pull_request when asked to monitor or babysit, or when work you were asked to do, such as a merge, still waits on checks or reviews", so "wait on CI" asks fordrive.
- pstack.
playbooks/shipping.md. Verifiers are Cursor cloud agents. PRs land bottom-up with--squash. The forge isghor Origin. It has six steps and no cleanup. Worktree and simulator cleanup is its own playbook. - orbit.
skills/playbook/playbooks/shipping.md, run by Ship by default or/orbit:ship./orbit:shipfirst runs Opening a PR when the session's todolist has no finished Opening a PR for the branch. Verifiers follow Independent verifier. The patch-id diffs against the PR's base, which is the parent's branch for a stacked PR. Under Landmain, steps 4 and 5 become a bottom-up fast-forward ofmain. gitops rebases each branch ontoorigin/main, the Check line stands in for CI, the patch-id is compared, and gitops pushesHEAD:main. The playbook index line drops the forge clause. Step 7 cleans up each branch that landed, after a successfulHEAD:mainpush or aMERGEDread-back. The session first leaves its own worktree withExitWorktreeandkeep, neverremoveordiscard_changes. Then gitops, from the session's restored directory, removes the task's worktree and every verifier and owner worktree made for it withgit worktree remove <absolute path>, never--force. A session that never calledEnterWorktreeskips the exit. When its working directory is the task worktree, as in a T3-bound thread, it keeps that worktree and its branch. A worktree with uncommitted changes stays. Then gitops deletes the local branch withgit branch -Donly when it holds nothing unlanded. Under Landmainthat is a fetch and a passinggit merge-base --is-ancestor <branch> origin/main, and gitops never switches, merges or fast-forwards the restored checkout. Under Landprit is a local tip equal to the PR'sheadRefOid, and step 4's--delete-branchalready removed the remote branch. The reply names each worktree and branch that stays./orbit:worktree-cleanupstill prunes abandoned worktrees. - Why. Cloud agents and Origin are not here. Under Land
mainno forge lands the run. A patch-id againstorigin/mainwould take in a stacked PR's parent and change once the parent lands. Under Landmainnothing pushed is rewritten, so no push needs force. pstack's cloud VMs leave nothing on the user's disk (inferred), and orbit's local worktrees do. The user asked "i have 140 worktrees?" when 126 worktrees held 196 GB and 124 of them had merged, then "branches too, merged prs left branches behind" (transcript audit, 2026-11-10). A squash merge leaves the branch's commits off trunk, so-drefuses them, and theMERGEDread-back is what makes-Dsafe. Ship by default runs Opening a PR first, but a bare/orbit:shipskipped it, and the user asked "why would u skip deslop?" (transcript audit, 2026-11-10). A session that ranEnterWorktreeis isolated in its worktree, and so is every subagent it spawns. Claude Code's worktree guard then refuses git that reaches another checkout throughcdor-C, "This session is isolated in the worktree, but this command redirects git to the shared checkout via -C. Refusing to run it".
ExitWorktreewithremoverefuses a worktree whose branch has a commit the session's starting commit lacks, and every landed branch has one. In real sessions 42 of 53 firstremovecalls were refused, and leads then passeddiscard_changes: truewithout the user. The ship cleanup eval at d840cd4 (tests/ship/cleanup.mjs --ref origin/main, 2026-11-19) reproduced both. While the lead was still isolated, gitops'scd <root> && git worktree remove ...andgit -C $R fetchwere refused, and the lead'sremovewas refused. AfterExitWorktreewithkeep, gitops removed the worktree and branch with no refusal. So step 7 exits withkeepfirst, and rules-shared.md'sgit -Cadvice ("One plain command per Bash call") names the isolation limit. In three runs of the same eval at 1e187da (2026-11-19), everygit branch -dfrom the restored directory was refused with "error: the branch 'fix/readme-typo' is not fully merged". TheHEAD:mainpush does not move the root checkout's localmain, and-dchecks a branch with no upstream against thatHEAD. The lead or gitops then rangit merge --ff-only origin/mainin the user's checkout to get past it. So under Landmaingitops deletes with-Doncegit merge-base --is-ancestorproves the branch is inorigin/main, the same shape as theheadRefOidrule, and leaves the user's checkout alone. A T3-bound thread works in its worktree without having calledEnterWorktree, soExitWorktreehas nothing to leave, git refuses to delete a branch checked out in a worktree, and removing the thread's working directory breaks every later shell call in it. So that thread keeps both (inferred). A session outside the worktree, such as a later/orbit:shipfrom the main checkout or the T3 merge-roundAgent(docs/harness/t3-code.mdstep 10), skips the exit, and gitops removes the worktree as usual.
- pstack. "A fork you could settle by running something → prototype it per
playbooks/prototype.mdinstead of asking." (poteto-mode/SKILL.mdNon-negotiables). de8ee15 classified the fork and said "Sketch it via the Prototype playbook (playbooks/prototype.md)", with the full-autonomy grant clause and "the Always-pause list in Autonomy". - orbit. The de8ee15 rule, grant clause included, in rules.md Non-negotiables (Trigger list). The sketch runs as
/orbit:playbook prototype, and the Always-pause list is the one in Task End. The trigger coversAskUserQuestionand a numbered "your call", "your pick" or "(Recommended)" choice in the reply. A fork no run settles goes toorbit:advisor, every open fork in one brief. The lead acts on the advisor's pick and reports it, with what the user could say to change it. It asks the user, with the pick marked recommended, only for an Always-pause item, a gate the user named, or a pick that drops or changes what the user asked for. A skill's own question, such as/orbit:setup's or Shipping's native-verifier consent, goes to the user as the skill says, and an advisor pick never counts as the user's yes to it. The Prototype playbook's step 2, "let the user pick directions before building", stays with the user on purpose. It is a taste call and that playbook's own step. Before it asks, the lead finishes the work that doesn't hang on the answer. An advisor that can't run on Fable retries on Opus, so an outage never sends a fork to the user. Under the grant, the advisor's pick is the default for a call only the operator can make. rules.md Writing the reply adds "Never end on a choice the advisor could make." - Why. Orbit has no Autonomy section, so its pause list sits in Task End (Task end). The lead asked the user to make calls as "Your call: 1. ... (recommended) 2. ..." without briefing the advisor. In 4 sessions the user pushed back with "did u check with advisor?", "ise advisor to decide dont ask me", "use advisor for decisions dont be blocked" or "consult" (transcript audit, 2026-11-20). The old trigger named only
AskUserQuestion, so a numbered choice in the reply slipped past it (inferred). pstack's Never Block on the Human says "Proceed, present the result, let the human course-correct." (poteto-mode/SKILL.md), and the advisor's pick is what the lead proceeds on. A skill's own question asks for the user's consent or knowledge, which the advisor cannot give. The native-verifier consent is the user's yes to a fallback that T3 forbids unasked (Independent verifier).
- pstack. Step 4 delegates to "a subagent using your configured feature model" and says "Commit liberally" (
playbooks/feature.md). Step 6 rebases into small commits. The closing paragraph says "That owner fans out internally after the blocking phase." - orbit. The same steps, with a fresh
orbit:poteto-agentper the Delegate gate and gitops committing each reviewed unit. Step 6 keeps the rebase, and gitops runs it (Commits). The single owner fans out oneorbit:poteto-agentper unit. Units the throughput checkpoint (step 3) marks independent, with disjoint files, run in parallel under the Parallel cap, each in its ownisolation: "worktree"worktree, and count tries as other units do. Under a T3 lead withdelegate_task, each parallel unit is a same-providerdelegate_taskchild instead. On every host the delegate commits its unit in its own worktree, and gitops cherry-picks each kept commit by SHA (Own-checkout children). Units that share a file run one at a time. Parent-level fan-out for independent artifacts runs as read-only agents outside the Delegate gate. - Why. pstack's step 3 says "Disjoint files, services, or layers parallelize. Shared writes serialize." Orbit's Delegate gate once ran every unit one at a time, and the user asked "why are u not using orbit why are u working without subagents" (transcript audit, 2026-11-10). A delegate works only in the path its brief gives, so a parallel unit needs its own worktree and a way back to the task branch. Artifact slices write no code, so the Delegate gate doesn't apply.
- pstack.
skills/swarm/SKILL.md. Workers run on theswarm workersmodel,claude-opus-5-5-high-fastby default, "withsubagent_type: generalPurpose,environment: "cloud",run_in_background: true".environment: "local"is for "access to something on the user's computer". It setsdisable-model-invocation: true. - orbit.
skills/swarm/SKILL.md, vendored at 0b46ec3 (skills.json). Workers run on thecodemodel, witheffortper rules.md. Each is ageneral-purposeAgentcall withisolation: "worktree"on a named branch or commit.isolationis omitted only when a worker needs what a fresh worktree lacks. Under a T3 lead withdelegate_task, a worker that writes is a same-providerdelegate_taskchild, and a read-only worker is a foregroundAgentwith no worktree (Own-checkout children). Workers spawn up to the Parallel cap in one message, and the next starts as one finishes. Every brief carries the Machine line. The skill runs through the Skill tool. - Why. The adapter maps the default slug to the
codemodel,cloudtoisolation: "worktree",localto omitting it, andcloud_base_branchto telling the worker to check out the ref. Every worktree is on the user's machine, so "the user's computer" becomes what a fresh worktree lacks (inferred). For the same reason, the worker count and what each worker opens on screen reach the user.
- pstack. Architect traces every system the change touches, sketches two shapes, and picks the one where "each piece of state has one owner, each task has one way to do it, and callers need to know least". "For a large or contested design, run the arena skill instead of sketching alone." (
skills/architect/SKILL.md) - orbit. The same text as
/orbit:architect, routing to/orbit:arena, whose two default runners apply (Panel seats). In an/orbit:playbookflow it reuses the flow's earlier/orbit:howoutput for the trace. The sketches stay in the lead's reply, and any stub code goes through the playbook's Delegate gate. Arena's runners areorbit:poteto-agent, each withisolation: "worktree"under the Delegate gate's parallel exception. Under a T3 lead withdelegate_task, each runner is a same-providerdelegate_taskchild instead (Own-checkout children). The judge is the roster seat under T3 Code, else anExploreagent on the first of thejudgmentandcodemodels that differs from the parent's, and arena skips it when neither differs (Arena judge). In a playbook flow, each runner commits its candidate in its own checkout, gitops cherry-picks the base's commit, and a freshorbit:poteto-agentfolds in the grafts. - Why. A
/orbit:howanswer from earlier in the flow already traces the affected systems, so a second trace repeats the read (inferred). Architect no longer names runners, a runner prompt or a rationale template, so the absolute-path and two-runner edits went with them. Under the Delegate gate the lead writes no code, so a sketch that turns into stub code needs a delegate. The adapter mapsreadonly: truetoExplore. Arena names nosubagent_type, so poteto-mode'spoteto-agentrule applies. The parent writes no code under the Delegate gate.
- pstack. "Never edit a skill under
plugins/cache/. Before approval, show its destination,.cursor/skills/<skill>/or~/.cursor/skills/<skill>/. After approval, copy the skill folder there and edit the copy. Never open a PR or push." Backlog has no destination named. (skills/reflect/SKILL.md) - orbit. Reads its transcript by
${CLAUDE_SESSION_ID}, lists Backlog as/orbit:correct <pattern>candidates, hands large edits and description tunes toskill-creatorif installed, and edits a plugin skill in its source repo, recording vendored changes inskills.jsonedits. - Why. Claude Code has no
.cursor/skills/override folder, and the plugin cache is overwritten on update. Orbit fixes its own skills in a PR (rules.md Non-negotiables, Broken orbit skill), so "Never open a PR" would clash.
- pstack. Cite the measurement in the PR (
playbooks/perf-issue.mdstep 5). - orbit. The measurement goes in the PR body when Land is
pr, else in the reply. - Why. Under Land
mainno PR opens.
- pstack. Each attempt goes to "a subagent using your configured hillclimb model", and parallel attempts each get their own worktree. "One commit per accepted fix, staging only the files you changed". An unattended run borrows "the wake mechanism from the Autonomous run playbook" (
playbooks/hillclimb.md). - orbit.
skills/playbook/playbooks/hillclimb.md. Each attempt is a freshorbit:poteto-agent, and gitops commits each accepted fix by path. Parallel attempts follow the Delegate gate's parallel exception, and gitops cherry-picks a kept one onto the task branch. The wake rule is inlined (Wake mechanism). - Why. Delegates never commit. A delegate works only in the path its brief gives, so an isolated attempt needs its own worktree and a way back to the task branch. Autonomous run is not ported.
- pstack. Under
/loop, "An event to watch (CI, a merge, a ref advancing) gets a watcher subagent that wakes you on the event", with a long heartbeat as fallback (playbooks/autonomous-run.mdstep 2). - orbit. Hillclimb step 5 checks each wake tool on its own, per the host adapter (T3 Code host), and says which it used. The loop and every heartbeat run on
schedule_taskbound to the thread when that tool is present, else under/loop. A PR event, CI, a review or a merge, getswatch_pull_requestwhen that tool is present, withuntilas Babysit's Host watcher says. In T3 with that tool off, a wait on a merge takes the heartbeat, one read of the PR per wake. Any other PR event there gets one read per turn and a handback, with noMonitor. Outside T3, a PR event gets aMonitorwatcher. Any other event, such as a ref advancing, gets aMonitorwatcher. When the run waits on thatMonitor's event, its task id goes tot3_background_waitwhen that tool is present. A background task getst3_background_waitwhen that tool is present, elseMonitoror its finish notification. An event watcher keeps a long time-based heartbeat as fallback. - Why. Orbit's own mapping, since pstack's adapter has no row for a watcher subagent. T3's session runtime instruction "Background work you wait on" says "To wait for a background command or Monitor you started (tests, a build), call
t3_background_waitwith its task ID before you end your turn." (system text T3 injected into a session, read 2026-11-17). That text is not in the c2eb8f3 repo snapshot. The tool's own description excludes "work left running (dev server, watcher)", so the rule names only aMonitorwhose event the run waits on, and a PR event keepswatch_pull_request, because T3 says "For PR waits, no sleep, loops, gh --watch, Monitor or ScheduleWakeup." (runtimeInstructions.ts:3in pingdotgg/t3code at c2eb8f3). A watch also wakes once when its PR merges or closes, so a merge needs no heartbeat while the watch tool is on (Babysit). With it off, a handback would stall an unattended run on a merge, so the merge keeps the heartbeat.Monitorwakes the session on the event, and no subagent has to poll. In T3,/loopwith no interval wakes throughScheduleWakeup, a timer inside the session (T3 probe lane, 2026-11-16).schedule_taskis T3's own scheduler, so the adapter prefers it there. A headless lane without T3 tools armedMonitorand added no/loopheartbeat, because the old sentence named/loopthe heartbeat tool only for the loop (T3-2 fix-round lane, 2026-11-16). Each tool is checked on its own because T3 can turn one off and leave the rest (T3 Code host).
- pstack.
playbooks/autopilot-full.md. Each owner is a Cursor cloud agent that opens its PR early, reports the code-ready head, pushes "each finished unit before you start the next", and merges once the root's swarm verifies it. "Assume the VM dies." Trails go in "the root agent's store". Each tick re-reads the playbook "from trunk". - orbit.
skills/playbook/playbooks/autopilot-full.md. Each owner is ageneral-purposeAgentcall withmodel: "opus"(thejudgmentmodel) andisolation: "worktree"that runs/orbit:start <item>and returns at merge-ready. Under a T3 lead withdelegate_task, only each owner's build round is a same-providerdelegate_taskchild. It never watches a PR, reads its PR once withwatch-pr --onceand ends with its PR, head and check state, and the root runs babysit on the PR. Until its wake, the tick judges it by its trail and git side effects, nevertask_status, and a stuck owner getstask_cancelbefore its replacement. Each later round, a fix, a rebase or the merge round, is a foregroundAgentin the root's thread that works in the owner's worktree (Own-checkout children). In T3 withoutdelegate_task, owners run one at a time, and the reply says the tick cannot fire while one runs. The owner reports code-ready as acode-ready <sha>trail row, and each tick starts the swarm at a new code-ready SHA. The PR opens in Opening a PR, after/orbit:deslopand/orbit:no-comments, not early alongside the work. "Assume the owner dies." Owners commit each finished unit before they start the next, instead of pushing it. Trails andagents.tsvgo in$TMPDIR/orbit-autopilot/<slug>/. Swarm lanes are swarm workers on thecodemodel, briefed as shipping step 1 briefs a verifier, with the Check line. They read the code-ready SHA from agit archive -oexport, so each lane is a foregroundAgentwith no worktree, and under Landprgitops posts the root's verdict line on the PR. Owners run in parallel under the Parallel cap. A fresh owner takes each later round, and its brief says to callEnterWorktreewithpathset to the prior worktree before any other work. The merge round, briefed also with the verifier worktree paths, runs only shipping steps 3, 4 and 7. Under Landmain, the merge is gitops pushingHEAD:main, then step 7. Ticks re-read the playbook without "from trunk". Under T3 Code with a roster, the root also spawns one read-only audit seat on another family per full swarm (Seats on another family). It is additive and never required for a clean verdict. Where babysit says to report a rebase, the owner has gitops rebase its own branch and push it with--force-with-lease=<branch>:<sha>, pinned to the SHA it last pushed or took over (Force-push). - Why. The adapter maps a cloud agent to
isolation: "worktree"andresumeto a fresh agent. T3 sends a separate checkout and a background agent todelegate_task, so in T3 an owner's build round is a child thread (Own-checkout children). A fresh agent starts pinned to the root's working directory and can't run git in the prior worktree. Later owner rounds inherited the root's worktree lock and wasted 3 Opus rounds (transcript audit, 2026-11-10). A pinned subagent may switch into a worktree under.claude/worktrees/of the same repo, soEnterWorktreewithpathgets it there. Step 7 removes the landed branch and the worktrees made for it (Shipping). A subagent reports once, when it returns, so the code-ready report goes to the trail the root reads. Orbit pushes nothing before Opening a PR, so a dead owner's work stays in its worktree. A subagent returns its verdict as text and gitops runs every forge write (Independent verifier). Full shipping would spawn a second verifier on a verified head. The playbook ships in the plugin, not the project's trunk. Thejudgmentmodel (Opus by default) for owners is orbit's choice, since each owner runs a whole flow. The lanes stay swarm workers, as pstack's "Verification is the exception and stays on theswarm workersmodel" says, and swarm workers run oncode. pstack'sswarm workersdefault was a grok slug at de8ee15, so its lanes brought another family into the verdict, and the audit seat keeps one there. e1cc031 moved that default toclaude-opus-5-5-high-fast. Lanes run the app and write scripts, so they stay on Claude (inferred). A child thread can stall or fail, so a clean verdict doesn't wait on it (inferred).
- pstack.
playbooks/autopilot-stack.md. "Run the Autopilot-full playbook, but do not merge." Owners "stop at verified" in place of step 5 and hand the PRs to the operator. A PR that builds on another is based on "the other PR's branch in place of trunk", each PR "rebases onto its parent PR's branch", and "A new head on the lower PR restacks the PRs above it". The root writes no topology of its own. - orbit.
skills/playbook/playbooks/autopilot-stack.md. It runsplaybooks/autopilot-full.md, named as a file. A dependent owner's brief names the parent's branch and its exact tip SHA as the base, in place oforigin/main, and the owner starts its own branch at that SHA before any work. - Why. Orbit names playbook files, as its other autopilot edits do. The adapter says "Name the branch or commit the worker checks out." Git won't check out a branch that another worktree holds, so the owner starts its own branch at the SHA.
- pstack.
playbooks/multi-phase-plan.md, five steps. Step 2 says "Ask the operator only about a product or preference call that no run can settle." Step 3 says "Write the plan file", with no location. Step 5 says "Hand back the plan path and stop. Execution starts on the operator's explicit go.", and names no execution playbook. It has no skeleton and no check script (2f87c85 deletedcheck-plan.mjsand its test). - orbit.
skills/playbook/playbooks/multi-phase-plan.md, with no gates. Under Landmaina PR in the plan is a task branch. Step 2 briefsorbit:advisoron a fork no run settles and asks the operator only where rules.md Non-negotiables says to (Test before asking). Unless the operator names a path, the plan goes under$TMPDIR/orbit-autopilot/<slug>/docs/. Step 5 adds that, per Ship by default, the go runsautopilot-full, orautopilot-stackwhen the operator said "don't ship" or "I'll land it". - Why. rules-shared.md Files keeps working notes out of the repo, and
$TMPDIR/orbit-autopilot/<slug>/is the autopilot store the plan's run uses. Ship by default grants landing authority unless the operator opts out (Ship by default), so orbit still names the run that executes the plan. The de8ee15 skeleton's lane placement table and control-skill paragraph clashed with T3 Code's thread and device rules, and the cut removes both. pstack's step 2 reads as a playbook's own question. rules.md Non-negotiables sends such a question to the user as the skill says, so the line kept a plan's forks from the advisor (advisor diff review, 2026-11-20).
- pstack. Bug reaches the surface "via the control skill (Non-negotiables)", refactoring "via the relevant control skill", and perf "via the matching control skill". Bug uses "Cursor's
/loopcommand". Investigation re-routes "to Bug fix or Feature", says "No PR, no babysit" and "see Autonomy". Prototype says "Route to Feature". - orbit. The control skill is the project's
verify-*skill when it has one, else/orbit:control-uior/orbit:control-cli. Bug uses/loop. Investigation and prototype route through/orbit:start <goal>. Investigation says "No commit, no PR, no babysit" and drops the Autonomy pointer. Delegate steps name a freshorbit:poteto-agentper the Delegate gate. The Worktree gate names hillclimb and figure-it-out runs by the commit type of what they change. - Why. A project's own verify skill knows its app. Orbit commits each unit as it passes review, so investigation rules out a commit too. Orbit has no Autonomy section. Branch names must be commit types.
- pstack. No waiver rule. Prototype covers "a throwaway sketch to make a design or behavioral decision cheaply" (
poteto-mode/SKILL.mdPlaybooks). - orbit.
startstep 2 keeps a step the user waives ("no commit", "no subagents", "skip e2e", "just testing", "quick check") in the todolist asskip: user waived. "Just testing" or "quick check" waives the commit, the PR and shipping. Step 1'sprototypeline adds brainstorming or iterating on UI the user isn't sure they'll keep. - Why. Orbit adds gates and ships by default, so a full flow runs where pstack's would stop at a PR. The user said "no need to put readme or commit this is just testing" and "when we are brainstorming ui dont go all the way u take forever for small changes" (transcript audit, 2026-11-10). The skip stays in the list, where the user can see it and overrule it.
- pstack. Route to
figure-it-outwhen no bundled playbook fits, or for large, cross-cutting or step-away work. figure-it-out's Start reads "the Principles section of the poteto-mode skill" (skills/figure-it-out/SKILL.md). e1cc031 deleted Orchestrate, so program-scale work routes to figure-it-out too. At de8ee15 a standing project-scale program "routes to Orchestrate instead". - orbit.
startroutes to/orbit:figure-it-outon the same triggers. Program-scale work routes to/orbit:figure-it-outtoo, as upstream, unless the goal is already a queue of PRs, which routes toautopilot-full. figure-it-out's Start reads the playbook's Principles, Stack rules and gates, and its last todo is Opening a PR when the run changes code. - Why. The gates keep a figure-it-out change on the same worktree, delegate and Opening a PR path as every other change.
- pstack. poteto-mode indexes Runtime forensics, Trace forensics, Visual parity, Eval, Authoring a skill and Autonomous run (
poteto-mode/SKILL.mdPlaybooks). de8ee15 also indexed Orchestrate. - orbit. None of them is vendored (the
playbook-playbooksskip list inskills.json).starthas no route for them, so such a task falls to/orbit:figure-it-outbecause no flow fits. Orbit removed Orchestrate on 2026-11-17 by operator decision, after upstream deleted it at e1cc031, and program-scale work routes to/orbit:figure-it-out. - Why. Orbit's choice. A playbook is adopted when a task needs it, as unvendored skills are (Vendored subset).
- pstack. e1cc031 has no such trigger. de8ee15 said "Agent-facing prose also follows the create-skill skill (Cursor's built-in for authoring SKILL.md files)." (
poteto-mode/SKILL.mdNon-negotiables) - orbit. Dropped from rules.md Non-negotiables.
- Why. It is Cursor's built-in and not in pstack's
skills/, so there is nothing to vendor.
- pstack. Pause safely, Session pickup and Worktree and simulator cleanup are poteto-mode playbooks. Pause safely writes its note to "a file like
/tmp/<slug>-resume.md". - orbit.
skills/pause-safely,skills/session-pickupandskills/worktree-cleanup, each with its ownskills.jsonentry. pause-safely has gitops commit thewip:checkpoint and writes the resume note to the memory fileresume-<slug>.md, one per paused task slug, as pstack names one per slug. session-pickup reads the note for the slug the user names, else the one whoseoriginSessionIdmatches the prior transcript. session-pickup keepsdisable-model-invocation. When a pause hands the thread back to the user, pause-safely lists the PRs the thread watches and callsunwatch_pull_requeston each, when those tools are present. Only the lead unwatches, and a subagent or delegated child skips that part of step 1. Before imminent compaction the lead keeps the watches. Either way the resume note names each watched PR and, under T3 Code, the thread'sthreadId, and says that a watched PR resumes through babysit. The reply names each PR unwatched. session-pickup sends an open PR whose watch a Stop or pause ended to babysit without a new ask. Babysit's Host watcher waits on it, and arms the watch while checks are pending. Pickup reads each PR's state withgh pr view <pr> --json state. When the note records the priorthreadId, pickup callslist_thread_pull_requestson that thread and skips a PR it still watches. It also skips a PR the trail unwatched at a handback other than a pause, and never callswatch_pull_requestitself. The adapter's Stop row holds both rules (T3 Code host). - Why. Each runs outside a task flow, so each needs its own command. A model-invocable pause-safely runs on "pause" or before compaction. rules-shared.md Files puts working notes in auto memory, not the repo, so only the folder differs from pstack. Orbit once wrote one shared
resume.md, and a pause in another session overwrote it ("the resume.md was overwritten what work do we have left here?", transcript audit, 2026-11-10). T3's runtime prompt says "Before handback, call unwatch_pull_request on any PR you watch." and "Call watch_pull_request when asked to monitor or babysit, or when work you were asked to do, such as a merge, still waits on checks or reviews, even after an earlier unwatch." (runtimeInstructions.ts:3in pingdotgg/t3code at c2eb8f3). A compaction keeps the session working, so it is no handback. T3's feature-lifecycles doc says "A Stop also holds the queue, ends the thread's pull request watches, and drops wakes its delegated tasks still owe, so nothing automatic restarts the thread." (docs/orchestration-v2/feature-lifecycles.md:254). So the PR work a Stop or pause cut off is work the user asked for that still waits, and resuming it is the ask. Babysit already waits on the PR per Host watcher, so pickup routes there instead of arming its own.list_thread_pull_requeststakes "Thread. Default: this thread." (pr-tools.ts:313in pingdotgg/t3code at c2eb8f3), and pickup runs in a new session, so a call with nothreadIdfinds none of the prior thread's PRs. The resume note records the priorthreadIdfor that call. A PR the prior thread still watches already wakes that thread, so pickup leaves it there (inferred). A PR the lead unwatched at an ordinary handback had finished its part, so it stays as the user left it (inferred).
- pstack. The audit reads transcripts from
~/.cursor/projects/<slug>/agent-transcriptsand fetches PRs through Origin's CLI, elsegh(poteto-mode/scripts/worktree-audit.sh). The pinned and active chats come "from the user or sidebar". Step 6 also clears Cursor's app state (poteto-mode/playbooks/worktree-cleanup.md). - orbit.
/orbit:worktree-cleanup. The audit reads~/.claude/projects/<slug>/**/*.jsonl. A worktree's newest session is one started in it or one of the repo's that names its path. A worktree Claude Code has locked ishold-lockedand never removed. PRs come fromghonly. The active sessions come from the user. The Subagents and Autonomy pointers and the Cursor app-state reclaimer are dropped. - Why. Claude Code stores sessions per start path and locks a live session's worktree. Origin's CLI and the sidebar don't exist here, and Claude Code has no Cursor app state.
- pstack. A changed run ships "at most one PR of proven corrections" (
skills/maintain-verification-skill/SKILL.md). - orbit. One change of proven corrections goes through Opening a PR. Clean or blocked runs land nothing.
- Why. Under Land
main, Opening a PR leaves an unpushed task branch, not a PR.
- pstack.
skills/ketchup/SKILL.mduses/bro's "voice, not its task" and draws charts with "a chart or diagram from/visualize". - orbit.
skills/ketchup, vendored at the pin withoutbro. The voice line is inlined. A chart for a trend or comparison goes as inline HTML throughhtml_preview, thenhtml_render, when both T3 tools are present, per the adapter's proof table row (T3 Code host). Without both,/visualizemaps to Mermaid charts (xychart-betaorpie). Diagrams stay Mermaid. Autopilot ticks and replies read it by path. - Why.
ketchupborrows onlybro's tone, so one sentence carries it. Orbit vendors no/visualize, and a Mermaid chart is text a reply can carry. T3's system prompt says "Answer data with a trend or comparison with a chart and show any visual the user asks for here: pass an HTML file path or inline HTML tohtml_preview, thenhtml_render." (orchestrationInstructions.ts:23in pingdotgg/t3code at c2eb8f3). Ketchup keeps its turn read-only, so the chart goes as inline HTML, not a file it writes.
- pstack. Flows drive the surface with "the matching control skill",
control-uiorcontrol-clifromcursor-team-kit.control-uicreates "before/after evidence forverify-this". - orbit.
skills/control-uiandskills/control-cli, vendored at the pin. The flows,rules.mdand the verifier brief fall back to them when a project has noverify-*skill.verify-thisbecomes theorbit:verifieragent. - Why. Without them, a project with no verify skill has no way to run a change as a user would. Orbit vendors no
verify-this, andorbit:verifierdoes its job.
- pstack. 58 skills (
pstack/skills/). - orbit. The subset in
skills.json: all 24 principle leaves,arena,ketchup, anddeslop,control-uiandcontrol-clifrom cursor-team-kit, among others. Every entry pins 0b46ec3. The playbook scripts come fromharness/pstack/claude-code/. They first pinned 1ba91fc there, when the 9580c23 scripts still carriedorch/and a test script that runs it. Upstream removedorchin c742dd6. - Why. Unvendored skills are adopted when a flow needs them.