Repository navigation
Conversation
|
Intent: Let operators select a harness, model, and reasoning effort for each crew member and retain those choices through delegated work, retries, and recovered conversations. Build on lifecycle PR #10686 by @CrysisDeu and preserve existing memory and permission boundaries. |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
Design Review (Opus 5.5) — 🟡 CONCERNSDesign-level review of Design-Verdict: CONCERNS The rule "skip installation model defaults when an explicit member backend is in a different namespace" is copied into four separate resolvers that can drift apart. Watch
[DESIGN-REVIEWED] 754eafc Description read[DESCRIPTION-READ] 165bc91555b2c8fdff1b18501fa262eb9b944d7ae424425f66331cd08ef5f2b2 That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
UX Review (Opus 5.5) — 🟡 CONCERNSUX-level review of UX-Verdict: CONCERNS The AI app picker lists "Use default: KAS (kiro-agent)" right beside "KAS (kiro-agent)", and the blind reader couldn't tell the two apart. Watch
Suggestions
[UX-REVIEWED] 754eafc Description read[DESCRIPTION-READ] ffb566fc52aeaa570136d5c6d170f4c7fcb1ba1959b1d422a8acc80635677261 That sha256 names everything this verdict read: the PR description, plus the 9 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
First Principles Review (Opus 5.5) — 🟡 CONCERNSPremise-level review of First-Principles-Verdict: CONCERNS One rider changes reset and restart behaviour for every member, including those that never set the new field, and its only evidence is a probe the author ran. Not justified as shipped
What this change shipsInventory (8 items) — 7 justifiedIntent: let each crew member run on its own AI app and model. This is an ADDITION.
[FIRST-PRINCIPLES-REVIEWED] 754eafc Description read[DESCRIPTION-READ] 83aebbe62e0e190eab79ec53df509c3ee5b86ff2b08c9a8454fe3680d82e697e That sha256 names the PR description this verdict read. Recompute it with |
Opus 5.5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
|
|
|
|
Dependency validation for #10686, authored by @CrysisDeu: the optional macOS run exposed two failures that reproduce on the unchanged lifecycle base
These are prerequisite-branch follow-ups, not regressions introduced by worker harness selection. They remain visible here because this PR deliberately depends on the lifecycle branch. The dependency and its author remain named at the top of the PR description. A third macOS failure, The fourth macOS failure was introduced here: a raw model-namespace literal violated the identity check. That comparison is repaired in |
|
68cdff8 to
e98203f
Compare
|
|
|
|
|
|
|
|
Use each member's backend when listing and validating its models and starting conversations. Capture delegated execution choices before queueing so retries and continuation retain the same settings. Build on the current SQLite member memory and execution context seams.
|
CI repairs in
Production and UI files are unchanged from |
|
|
|
|
|
|
|
|
|
[operator: bolichen97] Maintainer ruling: keep this PR as the one path for a per-member harness pin (#12451). #16200 is superseded by it. Options were (a) keep #11324, (b) keep #16200, or (c) merge parts of both. Both add the same Next, before merge (rebase needed, 24 files conflict):
|
|
Per the maintainer ruling, this PR is the path for the per-member harness pin. Please rebase onto current main and store the field through the harness vocabulary that #20359 added. Then the conflicts can be re-checked. While rebasing, please also bring over these parts of #16200, which this PR supersedes:
#16200 will be closed as superseded once this PR lands. Checked on head 754eafc: GitHub reports the branch as conflicting with main. A full review runs after the rebase. |
|
Status check on head 754eafc (stale-PR pass). I started the rebase onto current main. It conflicts in 45 files, and much of the touched code has moved (for example
Once (1) is decided, the rest of the ruling (harness vocabulary from #20359, one crew-aware |
bolichen97
left a comment
There was a problem hiding this comment.
[operator: bolichen97] My earlier change request is addressed on 754eafc. Approving to clear it; other open gates on this PR still apply.
|
I'm re-porting this onto current main as a new PR, scoped to crewmates: a per-crewmate harness pin ( |
CrysisDeu
left a comment
There was a problem hiding this comment.
Maintainer comment at 754eafcf. Zedmor, thank you for the design; the per-member pin, harness-scoped model/effort and the selector are the right shape and they survive. This head cannot land: a trial merge conflicts in 46 of 68 files because main split config/loader.py, subagent.py, chat_handlers.py and session.py into modules and deleted CrewEditorDialog.tsx, so it is a re-port, not a conflict fix, as @iamwhatever said on 10-23.
The re-port is #25485 (kyleseaman, scoped to crewmates, Zedmor co-authored). The 10-16 ruling stands and transfers to it: #11324's design is the per-member tier, #16200 is superseded. This PR closes once #25485 is through review. Decisions that bind #25485, recorded here so nobody has to dig:
- Blank vocabulary follows main (#20359): store through
backend_to_stored, read throughbackend_from_stored;""/absent = no choice,"kiro"= Kiro. This PR's""= explicit kiro-cli is out. - One crew-aware resolver for every
select_provider_backend(call site, plus a call-site guard test (main has 9 sites in 8 files). It must compose with #19599'sslot_backend(per-chat tier): precedencespawn> chat pick > member pin > member route > default. - Editor read path: (b), the owner-gated resolved-model readout;
acp_backendstays out ofGET /api/agents. - A crewmate's pin is validated at write time against
ACP_BACKENDS_MATE_WORKER_KIT(todaykas,claude), andcrewmate_worker_backendreads the member pin. A pin outside the kit is refused with a message naming the setting, notagent.member_acp_backend. codex/deepseek join the kit when someone lands a live-capture fixture (test/fixtures/acp_frames/<backend>/worker-kit-live.jsonl); kiro-cli cannot (reads--agentoff disk, no per-session channel). A member"kiro"pin keeps main's remap to KAS. - Missing or denied pinned backend fails closed with a visible refusal for enrolled members; no silent degrade to Kiro.
- Carry over from #16200: the changed-paths hot-reload rule and its
ConfigWatchtests, the call-site guard, the GET withholding, the warm-pool "unreadable config -> cold start" rule.
Zezhen's and Serena's review (iamwhatever's standing request) applies to #25485.
Posted by code-reviewer (Kiro Crew member) on behalf of @CrysisDeu. Full report on file.
|
@Zedmor this branch now conflicts with The maintainer ruling still stands: this PR is the path for the per-member harness pin. Could you rebase it onto current |
|
Thank you @Zedmor for this design. It is the shape that landed. The re-port, #25485, merged on 2026-10-30 as ffff62c, with you as co-author. It adds the per-crewmate harness pin ( |
Problem / Motivation
Goal: Let each crew member use its own AI app and model for conversations and delegated work.
A crew member can pin a model, but its harness is selected by installation-wide settings. A Kiro coordinator cannot assign individual members to Claude Code or Codex with their own model choices.
Why it matters
Operators can choose a harness and model for each role and retain that choice through delegated work and recovery. Cost savings depend on the selected models, workload and retries.
Not a goal
What changed (motivation → approach → change)
Both member editors offer an AI app choice and that app’s models. Use default follows existing routing; choosing Kiro explicitly remains distinct. Changing apps clears incompatible draft choices; returning to the opening app restores its model and effort. Catalog failures offer Retry without losing the draft. Template settings keep their own catalog.
Session and subagent routing selects the backend before the model. Queued-work reset successors and failed-compaction restarts retain the running app and model. Fresh allocations use current defaults. Delegated work captures choices before it can queue and keeps them through retries and continuation, including Auto and empty effort. Explicit caller choices take precedence; validation and effort receipts use the selected backend’s catalog.
Model readouts carry session and selection context: the composer follows its conversation, Settings uses normal-chat defaults, and member editors preview the member DM. Literal templates retain their own model and effort defaults. Profile saves update matching inactive DM slots so an older Profile-created model pin cannot survive the save. Profile supplies the expected app to the existing slot writes, which verify it under their mutation locks. Mismatched or unverified sessions remain unchanged.
This uses main’s SQLite member memory and canonical execution context. An enrolled member requires full saved-spec support; an unsupported app refuses startup before running work. Unenrolled members retain the app’s supported projections and tool/MCP channels. Profile settings follow the selected app while an open conversation keeps its app. The accepted crewmate profile settings and session template versus crewmate decisions define Profile placement and the distinction between a template and a crewmate.
config/sections.py,config-baseline.json,dashboard/handlers/agents.py,website/src/api/client/agents.tssession.py,session_allocation.py,subagent.py,subagent_manager/run.py,subagent_persistence.pydashboard/agent_admin/crew_records.py,dashboard/chat_handlers.py,providers/adapters/acp.ts,CrewProfileSettings.tsx,website/src/api/client/chatSlotSettings.tsMemberBackendField.tsx,useAvailableModels.ts,KiroCrewAgentsPage.tsx,en.manual.jsonand translated catalogsBackwards compatibility
Compatible: The optional member backend inherits current defaults when absent or null. An empty string explicitly selects Kiro. Durable run records without captured settings keep legacy behavior. Only a request changing the new backend field must also include model and effort choices. Readout context is optional; requests without it preview the member DM. Slot-write requests without the optional expected-backend guard retain their existing behavior.
An explicit member app in a different model namespace uses the member’s own model or Auto instead of installation template/global model defaults, even when an installation pin could be accepted by both apps. Inherited app choices keep the existing template/global fallback. Delegated role defaults apply only within the installation’s model namespace. Pin compatibility filtering remains separate from those inheritance rules.
Behavior change:
session_allocation.py:SessionAllocationServicecaptures the running backend for unpinned members too. Queued-work reset successors and failed-compaction restarts retain it afteragent.member_acp_backendchanges; a fresh allocation follows the new default.test_member_execution_selection.pycovers this through live reuse, real queued-work reset, failed-compaction restart and a fresh allocation. A portable probe on clean7bb9d576forwards the existing requested model to the changed backend after queued reset; this branch preserves the original pair. The probe uses a provider double and a synthetic model, so it demonstrates forwarding consistency rather than native model acceptance.Removes nothing: Existing memory, templates, backend choices and conversation storage remain in place.
Tests
Current revision
754eafcfca3fchanges only two test files relative to6e0bedb84265. CI frontend shard 3 found the old one-argument spy expectation inwebsite/src/test/AcpAdapter.defaults.test.tsafter the API gained optional routing context (11,745 pass / one fail / two skipped). The exact file reproduces eight passes and that failure before correction. Three forwarding cases now cover absent context, normal chat and explicit templates. All 50 neighboring adapter tests pass; the three cases also pass 20 serial runs, 10 parallel runs with neighbors and a shuffled run (660 executions).Linux and Windows backend shard 3 both found that
test/test_slot_running_readers.pylacked the new Profile guard in its reservation-reader inventory. Linux had 20,484 passes and that one failure. The inventory now explicitly declares the guard and credits its callers through the helper map; the source scanner and equality assertion remain intact. The exact file reproduces one pass and one failure before correction. The corrected check passes 20 serial runs, 10 parallel runs beside related reservation tests, and a shuffled run (92 executions). Formatting, import ordering, lint and the flake-pattern ratchet pass.Production code, owning docs and the base are unchanged, so the source-level validation below remains applicable. The periodic mechanism review retained the production design and confirmed earlier obsolete mechanisms remain absent; served reviewer model identity is unverified. A repeated related-gate attempt selected the full feature through the sibling commits' merge-base and was stopped as duplicate work; it is not counted as a pass. The focused correction checks above and the earlier full-floor limitations below are the recorded local evidence.
On
6e0bedb84265, rebased onto64e1cbb9dfcfafter #11780, all 641 selected backend interaction tests and all 1,040 tests in the two API coverage files pass. The backend selection exercises handler/subagent imports, member routing, canonical context, Profile slot guards, restart/continuation, and the changed main recovery paths. A scan of all 6,151 tracked Python files finds none of the ten retired identifiers from the compatibility review and no tokenization errors. The canonical memory files remain unchanged from this base. Independent source review clears this exact commit; served reviewer model identity is unverified.26b45b0ad7f1ultimately passed 93 CI checks.The broader frontend run passes 12,089 tests and fails only the generic
agentResolvedModelprobe inwebsite/src/test/ApiClient.coverage.test.tsx, which supplied positional text for the new context object. Its argument table now supplies a valid session key and selection kind; assertions are unchanged. The corrected case passes 20 serial runs, 10 parallel runs beside its neighbors (1,039 tests per run before the rebase), and a shuffled run. That full shuffled run has two failures in unchanged error-journal cases: absolute versus relative endpoint URLs. Removing only this fixture correction reproduces both failures plus the original invalid-context failure. Those two order-dependent failures remain documented, while the post-rebase ordinary run passes all 1,040 tests.Broader local gates are not wholly green. The related backend gate stops in its first batch after 13,713 passes, 70 skips and six template-skills ownership refusals. The exact six tests also fail on clean
7bb9d576in a linked worktree, and pass in ordinary clones of both7bb9d576and64e1cbb9. Their tests, conftests and ownership code are unchanged across the feature and those bases. The production ownership guard remains intact; later backend batches did not run. Electron has 2,579 passes, four skips and the same single process-tree failure reproduced on clean7bb9d576; its test, preload and implementation remain byte-identical to that proof. This host's process-list command omits the requested parent-PID column. Earlier broader runs also reproduced an unchanged Sage mock-task leak and four representative installer/grep host failures on that clean base. These limitations are not counted as passes.The previous CI boot-isolation assertion for
atomic_write._default_modealso reproduced on clean base (14 pass / 1 fail), versus 15 passing locally on the feature branch. The audit omits a lazy process-umask cache; the exact initializing call is unproven. One targeted CI retry passed all 63 integration tests on unchanged26b45b0ad7f1. The original intermittent failure remains recorded on the PR and does not waive fresh CI.Manual verification
An isolated gateway and Chromium pass 13 checks covering both member editors, scoped catalogs, default versus explicit selection, save/reopen, and restoration of the opening model and effort. Controlled 503 responses exercise both catalog errors in both editors. Retry reloads both lists, clears the alerts, and preserves the unsaved model and effort; subsequent real saves return 200 and persist those choices.
A Profile save for a conversation using another app shows the dedicated neutral confirmation. The guarded slot request refuses the mismatch without changing that conversation. Backend tests cover matching cold slots, unknown providers, allocation races, busy turns, children, and explicit Kiro. The earlier screenshot run used controlled catalogs and slot-capability responses.
The current production build also passes 15 browser checks through real Profile and slot endpoints with a simulated provider. A prior Profile model pin is replaced while the slot is inactive; both its next turn and Fresh start use the new model. Effort reaches the slot and provider-factory override, but the fake provider does not advertise effort support. Six further browser checks establish normal-chat composer and Settings readouts against a different member-DM default; readout endpoint responses are not intercepted. All temporary gateways and homes were removed.
Native Kiro and Claude each completed a delegated member task using Auto. Native Codex reached a sign-in refusal in two bounded attempts; inference on Codex is unverified. All harness gateways, helper processes and temporary homes were removed, and production provider settings remained unchanged.
Screenshots / video
Use default follows the installation's routing; selecting a named app pins that choice for the member:
The Profile confirms that the model choice was saved for new sessions. The open conversation keeps its existing app and selection:
Both catalog requests fail while a model/effort draft is unsaved. The lower readout shows the stored settings until Save changes; the editable controls retain the pending choices:
Recovery, scoped models, and the shared editor
Retry restores both catalogs and clears the alerts while retaining the same unsaved choices:
Related Issues
Addresses the compatibility review following #11780. The former lifecycle stack has landed; this PR targets main.
Checklist