0.28.1 — a limit stop is not a wait

.claude-plugin/marketplace.json

@@ -12,7 +12,7 @@
       "displayName": "Agent Stack",
       "source": "./plugins/agent-stack",
       "description": "Four skills: agent-orchestrator — tool-calling loops, pipelines with checkpoints, provider routing with fallback, memory architecture, plus the wallet side of reselling LLM access; agent-evals — run/trace/thread evals, LLM judges, and fixtures grown from production; agent-interop — MCP servers and clients, A2A agent cards, the MCP Registry, and gateways; agent-harness — system prompts, tool shaping, workflow-vs-agent, and auditing an agent system.",
-      "version": "0.27.2",
+      "version": "0.28.1",
       "author": {
         "name": "ssheleg",
         "url": "https://x.com/sshlg93"

CHANGELOG.md

@@ -1,3 +1,32 @@
+## 0.28.1 — a limit stop is not a wait
+
+There is no 0.28.0 package. The tag `v0.28.0` was pushed by mistake onto `cb8f699` (the 0.27.1
+source) before this change merged. Its release run (38332687325) was cancelled before the
+release and publish jobs ran, so nothing was published under that version. The tag ruleset
+does not allow deleting the tag, so it stays.
+
+- **agent-orchestrator §4 and `references/runtime.md` → *A limit stop is not a wait*.** A worker
+  is a workflow lane, a sub-agent or a session. When it stops on a provider usage or session
+  limit ("resets 5am"), it is restarted at once from where it stopped and is never left until
+  the stated reset. The same holds when it fails on a rate-limit error after the host's own
+  retries, or goes silent for 15–20 minutes with its task unfinished and no command of its own
+  running. How each worker is restarted:
+  - a lane is resumed with the same run id and the same arguments, and finished agents replay
+    from the cache;
+  - a sub-agent gets `resume` at its id;
+  - a session is re-prompted.
+
+  A second immediate stop backs off 2, 5, 10, then every 15 minutes while other work goes on.
+  Spend, credit and billing limits are a person's gate: they are reported once and never looped.
+  Every restart is logged.
+
+  The reset time in a limit message describes one account at the moment of refusal. An account
+  rotator or a short rate window usually makes it stale within minutes. A team running parallel
+  workflow lanes observed this: six stopped on stated resets hours or days away, and each ran
+  when restarted minutes later.
+
+  The references index row for `runtime.md` names the usage-limit stop.
+
 ## 0.27.2 — owner board and release ancestry checks
 
 - The repository validator now rejects missing boards, shifted rows, duplicate IDs

SKILL-CARD.md

@@ -5,7 +5,7 @@
 | Field | Value |
 |---|---|
 | Pack | `agent-stack` |
-| Version | `0.27.2` |
+| Version | `0.28.1` |
 | Skills | `agent-orchestrator`, `agent-evals`, `agent-interop`, `agent-harness` |
 | License | MIT |
 | Source | https://github.com/ssheleg/agent-stack |

docs/evidence/verification.md

@@ -10,6 +10,18 @@ This file exists because its absence read as zero exposure. `sshlg-skills` board
 
 ---
 
+## Release candidate — v0.28.1 (2026-10-13)
+
+The change is doctrine only: no script, validator or installer change. It was measured on the
+branch `feat/limit-stop-restart` (rebased on `e2120c2`) before any tag; nothing below is a
+published-artifact claim. `v0.28.0` is a mis-pushed tag on `cb8f699`, and its run 38332687325
+was cancelled before release and publish (`npm view` → 0.27.2 afterwards); see CHANGELOG 0.28.1.
+
+| REQ | What ships | How it was confirmed | Confirmed |
+|---|---|---|---|
+| Limit stop | The agent-orchestrator `SKILL.md` §4 bullet *A limit stop is not a wait* routes to `references/runtime.md` → *A limit stop is not a wait*. That section holds the three triggers, the restart table, the backoff, the billing exception and the restart log | `grep -n "A limit stop is not a wait"` on both files; the runtime contents list carries the section | **observed**, local |
+| Gate | `npm test` | EXIT=0 on 2026-10-13: `OK: agent-stack structurally valid (17 checks = 13 named + 4 per-skill, 4 skill(s), v0.28.1)`, `PASS: plant_guard — 9 cases`, `PASS: reap-own — 13 cases`, `PASS: installer — 11 case(s)`, all 25 `test/audit_regressions/*.py` exit 0; `claude plugin validate --strict` passed for `plugins/agent-stack` and the marketplace | **observed**, local |
+
 ## Released — v0.27.2 (2026-10-13)
 
 Source integration: PR47 merged at `053c0f8725d28e464c98d063e87618c15174849a`.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ssheleg/agent-stack",
-  "version": "0.27.2",
+  "version": "0.28.1",
   "scripts": {
     "test": "python3 test/validate.py && python3 test/plant_guard_test.py && python3 test/member_owner_gates_test.py && python3 test/reap_own_test.py && node test/installer_test.js && npm run test:audit",
     "test:audit": "for t in test/audit_regressions/*.py; do python3 \"$t\" || exit 1; done"

plugins/agent-stack/.claude-plugin/plugin.json

@@ -3,7 +3,7 @@
   "name": "agent-stack",
   "displayName": "Agent Stack",
   "description": "Four skills: agent-orchestrator — tool-calling loops, pipelines with checkpoints, provider routing with fallback, memory architecture, plus the wallet side of reselling LLM access; agent-evals — run/trace/thread evals, LLM judges, and fixtures grown from production; agent-interop — MCP servers and clients, A2A agent cards, the MCP Registry, and gateways; agent-harness — system prompts, tool shaping, workflow-vs-agent, and auditing an agent system.",
-  "version": "0.27.2",
+  "version": "0.28.1",
   "author": {
     "name": "ssheleg",
     "url": "https://x.com/sshlg93"

plugins/agent-stack/skills/agent-orchestrator/SKILL.md

@@ -134,6 +134,11 @@ Three decisions the rest follows from:
   passes every check that only looks for an exception.
 - **Cap the attempts and return the best partial**, because the alternative to a partial
   answer is not a better answer, it is no answer and a spent budget.
+- **A limit stop is not a wait.** A worker halted by a provider's usage limit
+  ("resets at 9pm"), a rate-limit error, or silence with its task unfinished is restarted
+  from where it stopped right away, not at the stated reset; back off only when the
+  restart fails again. Spend and billing limits are the exception — a person's gate.
+  Mechanics: [`references/runtime.md`](references/runtime.md) → *A limit stop is not a wait*.
 
 The hierarchy, the loop and the per-domain validators:
 [`references/patterns.md`](references/patterns.md).
@@ -346,7 +351,7 @@ there; this table is only an index.
 | [`references/memory-landscape.md`](references/memory-landscape.md) | **build or adopt** a memory layer, and what is settled practice |
 | [`references/context-engineering.md`](references/context-engineering.md) | the loop is **running out of window** |
 | [`references/kv-cache.md`](references/kv-cache.md) | you are deciding **what goes where in a request** — the static prefix against the trajectory, what a miss costs, the 2^N cache populations a runtime condition in front of the boundary creates, the named invalidators, and the four common optimisations that pay cache for nothing |
-| [`references/runtime.md`](references/runtime.md) | the agent must **survive a crash, a pause, a second message or a schedule** |
+| [`references/runtime.md`](references/runtime.md) | the agent must **survive a crash, a pause, a second message, a schedule or a usage-limit stop** |
 | [`references/governance.md`](references/governance.md) | the question is **permission, not cost** — what it may do, and how you prove it |
 | [`references/llm-proxy-billing.md`](references/llm-proxy-billing.md) | the product **resells LLM access** |
 | [`references/provider-lifecycle.md`](references/provider-lifecycle.md) | the question is the **workforce, not the loop** — provider lifecycle, canary trust, fleet budgets |

plugins/agent-stack/skills/agent-orchestrator/references/runtime.md

@@ -13,6 +13,7 @@ has only built the first one discovers the second in production.
 ## Contents
 
 - Durability, and where our own asymmetry was
+- A limit stop is not a wait
 - The interrupt/resume contract
 - Double-texting: four policies
 - Streaming that survives a dropped connection
@@ -42,6 +43,46 @@ Two properties earn their cost:
 - **A pause frees the worker.** An agent waiting for a human should hold no process. If
   waiting costs a worker, long approvals are quietly expensive and teams stop using them.
 
+## A limit stop is not a wait
+
+A coding-agent host shows a usage stop as *You've hit your session limit · resets 5am*.
+The time in that message describes **one account**, read when the request was refused. In
+practice an orchestrator that restarts the worker minutes later usually finds it runs. An
+account rotator may already have moved the session to an account with room, or the
+limit was a short rate window. Waiting hours for the stated reset idles a run that would
+have gone on.
+
+**Three triggers** send the orchestrator to restart a worker — a lane of a workflow, a
+sub-agent, a whole session:
+
+- it stopped on a usage or session limit, whatever reset time it names;
+- it failed on a rate-limit error after the host's own retries;
+- it went silent: no transcript or event write for 15–20 minutes, its task unfinished,
+  **and no command of its own still running**. A build or a test suite that runs for half
+  an hour is not a stall, and stopping that lane loses the step in flight.
+
+**Restart at once, from where it stopped**, never from scratch:
+
+| Worker | Restart |
+|---|---|
+| A workflow lane | stop the workflow, then resume it with the same run id and **the same arguments**: finished agents replay from the cache, and only the failed or stalled ones run again |
+| A sub-agent | send `resume` to its id, so it keeps its own context |
+| A session | re-prompt it: *continue from where you stopped* |
+
+Changed arguments invalidate the cache, so the whole run is paid for again.
+
+**Back off only on a second stop.** If the restart halts again at once, retry at 2, 5,
+10 and then every 15 minutes, and keep doing independent work in between. Do not sleep
+until the reset time.
+
+**Not every limit is retried.** A spend cap, exhausted credits or a billing error does not
+clear on a timer: someone has to pay or raise it. Report it to that person once and do
+not loop.
+
+**Log every restart** in the run's journal: which worker stopped, on what, when, and the
+attempt number. Ten restarts of one lane in an hour is a finding about that lane, not
+noise.
+
 ## The interrupt/resume contract
 
 The body has two mechanisms for one idea: `ask_user` raises a clarification error, and a