Skip to content

feat: select a harness and model per crew member - #11324

Closed
Zedmor wants to merge 1 commit into
mainfrom
feat/member-model-backend
Closed

Zedmor wants to merge 1 commit into
mainfrom
feat/member-model-backend

Conversation

@Zedmor

@Zedmor Zedmor commented Sep 16, 2026 •

Copy link
Copy Markdown
Contributor

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

  • Add a model provider, authentication path, or memory store.
  • Move an existing conversation to another AI app.

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.

Area Owning files
Member configuration and API config/sections.py, config-baseline.json, dashboard/handlers/agents.py, website/src/api/client/agents.ts
Conversations and delegated recovery session.py, session_allocation.py, subagent.py, subagent_manager/run.py, subagent_persistence.py
Readouts and Profile saves dashboard/agent_admin/crew_records.py, dashboard/chat_handlers.py, providers/adapters/acp.ts, CrewProfileSettings.tsx, website/src/api/client/chatSlotSettings.ts
Member controls and catalogs MemberBackendField.tsx, useAvailableModels.ts, KiroCrewAgentsPage.tsx, en.manual.json and translated catalogs

Backwards 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:SessionAllocationService captures the running backend for unpinned members too. Queued-work reset successors and failed-compaction restarts retain it after agent.member_acp_backend changes; a fresh allocation follows the new default. test_member_execution_selection.py covers this through live reuse, real queued-work reset, failed-compaction restart and a fresh allocation. A portable probe on clean 7bb9d576 forwards 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 754eafcfca3f changes only two test files relative to 6e0bedb84265. CI frontend shard 3 found the old one-argument spy expectation in website/src/test/AcpAdapter.defaults.test.ts after 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.py lacked 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 onto 64e1cbb9dfcf after #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.

  • Current readout checks: 192 focused backend tests, 12 canonical-selection tests and 207 frontend tests pass. Current Profile checks: 275 related backend tests pass, including busy sessions, sibling aliases, identity changes and reset interleavings. These selections overlap; they are not an aggregate test count.
  • Determinism: the readout's 11 backend and seven frontend regressions pass 20 serial runs, 10 parallel runs and shuffled runs. All 54 Profile guard/boundary cases pass the same schedule (1,718 executions); unchanged Profile frontend cases pass that schedule too (899 executions). The first guard implementation fails 12 boundary assertions, while the correction and independent verifier pass all 16 boundary cases. Restoring the old readout endpoint fails six routing cases; restoring the old identity or binding resolver fails three template-pin cases each.
  • Earlier feature validation, before these follow-up fixes: 339 member-selection/backend tests, 331 canonical-context/memory tests, 221 continuation/model-scope tests, and 260 frontend tests pass. The original 136 new backend cases pass 20 serial runs, 10 parallel runs and a shuffle; eight captured-restart cases pass 20 serial and 10 parallel runs; the ten continuation/namespace follow-up cases pass 20 serial, 10 parallel and a shuffled run (310 executions). Earlier published head 26b45b0ad7f1 ultimately passed 93 CI checks.
  • Black, subprocess encoding, isort, flake8, mypy, documentation, harness/memory boundaries, the flake-pattern ratchet, pinned Semgrep, TypeScript, frontend lint/localization, production build and bundle checks pass. The isolated template-lint tool also passes after explicitly selecting the supported Python interpreter. Handler/subagent imports and retired-symbol checks are reported with the post-rebase result above.

The broader frontend run passes 12,089 tests and fails only the generic agentResolvedModel probe in website/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 7bb9d576 in a linked worktree, and pass in ordinary clones of both 7bb9d576 and 64e1cbb9. 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 clean 7bb9d576; 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_mode also 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 unchanged 26b45b0ad7f1. 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:

Default versus an explicit AI app

The Profile confirms that the model choice was saved for new sessions. The open conversation keeps its existing app and selection:

Profile confirms the save for new sessions

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:

Catalog failures with retained draft and Retry

Recovery, scoped models, and the shared editor

Retry restores both catalogs and clears the alerts while retaining the same unsaved choices:

Catalogs recovered without losing the draft

Models scoped to the selected Codex app

Shared editor after saving and reopening default selection

Related Issues

Addresses the compatibility review following #11780. The former lifecycle stack has landed; this PR targets main.

Checklist

  • At most two commits (one is the norm), with a Conventional Commits title
  • New functionality and directly related regressions pass; broad-gate limitation documented above
  • Self-review completed; code follows project style guidelines
  • Documentation updated
  • No secrets, credentials, or internal references in the diff

@Zedmor
Zedmor requested review from a team and buluoray and removed request for a team September 16, 2026 15:21
@github-actions github-actions Bot added the readiness: checking Automated validation is still running label Sep 16, 2026
@Zedmor

Zedmor commented Sep 16, 2026

Copy link
Copy Markdown
Contributor Author

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.
Not a goal: Add a private-memory authentication path for Codex, weaken sandbox isolation, or make every harness capability interchangeable.

@github-actions

github-actions Bot commented Sep 16, 2026 •

Copy link
Copy Markdown
Contributor

GPT 6.1 Review — ✅ no blocking findings

GPT 6.1 completed its review of 754eafcfca3f3ab3beb09d24fb5d428896096ca2 and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 754eafc

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 754eafcfca3f3ab3beb09d24fb5d428896096ca2: <one-sentence reason>

@github-actions

github-actions Bot commented Sep 16, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Opus 5.5) — 🟡 CONCERNS

Design-level review of 754eafcfca3f3ab3beb09d24fb5d428896096ca2 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

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

  • Inheritance rule copied in four places. The description says "An explicit member app in a different model namespace uses the member's own model or Auto instead of installation template/global model defaults." Each site writes that check out for itself: config/loader.py:4439, config/loader.py:7410, session.py:855 (_session_model) and subagent.py:1395 (_capture_member_settings, which also has its own effort fallback chain). If a later change updates one site and misses another, a member's DM, a delegated run and the dashboard readout will each resolve a different model or effort for the same member. The likely symptom is that the readout shows one model while the delegated run starts on another. Nothing pins all the paths to the same answer for a given member and backend.
    Clears when: all four sites call one helper on KiroCrewConfig, for example member_inherits_installation_defaults(crew), or a single parametrized test checks that chat, subagent capture and readout give the same model and effort for every combination of backend pin (null, "" or another backend) and member model pin (set or unset).

[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 .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Sep 16, 2026 •

Copy link
Copy Markdown
Contributor

UX Review (Opus 5.5) — 🟡 CONCERNS

UX-level review of 754eafcfca3f3ab3beb09d24fb5d428896096ca2 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

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

  • Default and pinned rows look like duplicates. inherited_ai_app ("Use default: {{backend}}") puts the default's name into its label, so shot-01 shows the same app twice. The reader said they "seem to be the same app listed twice… no idea how they differ… no". This is the one distinction the field exists to make, and every member editor shows it. Fix: rename the row to say what it does, e.g. "Follow installation default (currently KAS)", and add a short line explaining what choosing an app does.
  • The Profile note talks about a setting the Profile panel doesn't have. In shot-02, "Saved for new sessions. This conversation keeps its current AI app." shows in a panel with no AI app field. The reader saw "no AI app setting… only Model". Users will look for a control that isn't there. Fix: "Saved. New conversations use this model; this one keeps its current setup."
  • Retry looks like a text box. In shot-03, <Btn> sits inside the field's flex-col wrapper, which stretches it to the full width of the field. The reader thought "it looks like a text box, not a button". That's the only way out of a failed catalog load. Fix: give it self-start (and an icon) so it looks like the other buttons.
  • The label style doesn't match its neighbours. "AI app" is sentence-case muted text, while the fields below are labelled "MODEL" and "REASONING EFFORT" (shot-03). The new field looks like it belongs to a different form. Fix: use the same label style as ModelField.

Suggestions

  • ai_app_hint: "pins it for this member" and "installation's AI app" were both unclear to the reader. Suggested wording: "New conversations with this crewmate use this app. Open conversations keep theirs."
  • Summary box: it hides only when the AI app is changed. When only the model is changed (shots 03/04), it still shows the old saved values. The reader asked which one is true. Either hide it or label it "Saved:" whenever the draft is changed.

[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 .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Sep 16, 2026 •

Copy link
Copy Markdown
Contributor

First Principles Review (Opus 5.5) — 🟡 CONCERNS

Premise-level review of 754eafcfca3f3ab3beb09d24fb5d428896096ca2 — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

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

  • Item 6 — rides along, inherited: the description says the allocation service "captures the running backend for unpinned members too". The only evidence offered is a probe "on clean 7bb9d576" that, by the author's account, "demonstrates forwarding consistency rather than native model acceptance". No issue is linked, and no end-user failure is shown.
    Subtraction: limit the captured backend in session_allocation.py to members with a non-null agents.<name>.acp_backend, so unpinned members keep base behaviour.
    Clears when: a linked issue or a test that fails on base shows a successor after a queued reset failing on the new backend with the forwarded model.

What this change ships

Inventory (8 items) — 7 justified

Intent: let each crew member run on its own AI app and model. This is an ADDITION.

  1. Both member editors offer an AI app choice and that app's models — justified
  2. New nullable config key agents.*.acp_backend; null inherits, "" means Kiro — justified
  3. Conversations and delegated work route to the member's app before choosing a model — justified
  4. Delegated runs persist model and effort through retries and continuation — justified
  5. An explicit foreign-app member skips the installation's model defaults — justified
  6. Successors after queued reset or failed compaction keep the running app, even for unpinned members — rides along, inherited
  7. Composer, Settings and editor model readouts follow session and selection context — justified
  8. Profile saves pass an expected app to slot writes and update inactive DM slots — justified

[FIRST-PRINCIPLES-REVIEWED] 754eafc

Description read

[DESCRIPTION-READ] 83aebbe62e0e190eab79ec53df509c3ee5b86ff2b08c9a8454fe3680d82e697e

That sha256 names the PR description this verdict read. Recompute it with .github/scripts/pr-description-capture.sh; if it differs, the description changed after this verdict was formed and any finding drawn from the description is unproven.

@github-actions

github-actions Bot commented Sep 16, 2026 •

Copy link
Copy Markdown
Contributor

Opus 5.5 Review — ✅ no blocking findings

Reviewed 754eafcfca3f3ab3beb09d24fb5d428896096ca2 — this comment is updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] 754eafc

Verdict parsed from the review's SHA-scoped output markers for commit 754eafcfca3f3ab3beb09d24fb5d428896096ca2.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 754eafcfca3f3ab3beb09d24fb5d428896096ca2: <one-sentence reason>

@github-actions github-actions Bot added readiness: passed Eligible automated validation passed for the current revision and removed readiness: checking Automated validation is still running labels Sep 16, 2026
@Zedmor

Zedmor commented Sep 16, 2026 •

Copy link
Copy Markdown
Contributor Author
  • Runtime integration scope span=0c42ab1c3c62

Disposition: rebutted as a scope recommendation. The accepted task includes live delegated work on Kiro CLI, Claude Code and Codex on this clone; delivering only the configuration selector would leave that acceptance incomplete. These fixes address failures reached by that live path.
The Codex adaptation requires both the MCP-approval marker and trusted, correlated server/tool parameters. It translates wire identity before the existing permission checks; it adds no authentication path or grant, and private Codex members remain refused. The negative cases are exercised by test_codex_mcp_permission_uses_verified_identity_not_shell_kind.
The platform changes retain the existing atomic no-replace contract and repair parent-PID observation. Splitting them would add prerequisites to the same acceptance path. Keeping this bounded integration in one PR is an intentional tradeoff; this does not claim that separate PRs would be invalid.

@Zedmor

Zedmor commented Sep 16, 2026 •

Copy link
Copy Markdown
Contributor Author
  • Platform prerequisites for live member execution span=c84bf75433ec

Disposition: rebutted as a scope subtraction. Private-member provisioning requires atomic create-if-absent publication, and runtime identity uses the parent map. Removing these changes breaks the requested live member validation on the tested Linux host.
The failure is reproducible with glibc 2.26 and procps-ng 3.3.10: libc omits renameat2 despite kernel support; ps -Ao pid=,ppid= returns only the PID column, while ps -A -o pid= -o ppid= returns PID and PPID. This is a concrete runtime failure, independent of the unrelated Node installer fixture failures.
The rename fallback invokes the same kernel no-replace operation on supported 64-bit ABIs; unsupported cases still refuse. Native publication/collision-refusal and real parent-map tests pass locally. The explicit missing-libc-symbol regression is included in e98203f. The completed macOS run reported no platform-helper failures; its other failures and baseline evidence are documented at #11324 (comment).

@Zedmor

Zedmor commented Sep 16, 2026 •

Copy link
Copy Markdown
Contributor Author
  • Codex MCP identity is required by delegated tool use span=7d1cd202167f

Disposition: rebutted as a scope subtraction. Installation-wide Codex selectability did not prove that delegated MCP calls passed the existing permission gate. The requested acceptance includes functioning Codex workers, and the adapter defect was reached during that validation.
Codex represents MCP approvals with kind=execute. Treating the correlated MCP request as a shell call sends it to the wrong policy path. adapt_mcp_permission translates only the Codex-specific approval marker plus trusted cached server/tool identity; uncorrelated, marker-free, and other-harness events retain their previous classification.
This adapter fix and its negative regressions stay with the working mixed-harness feature. It does not authorize private-memory access: the sandbox credential mask, shared-broker exclusion, and private Codex refusal are preserved.

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention readiness: passed Eligible automated validation passed for the current revision and removed readiness: passed Eligible automated validation passed for the current revision readiness: action required A blocking check or review needs attention labels Sep 16, 2026
@Zedmor

Zedmor commented Sep 16, 2026 •

Copy link
Copy Markdown
Contributor Author

Dependency validation for #10686, authored by @CrysisDeu: the optional macOS run exposed two failures that reproduce on the unchanged lifecycle base 547e1ddfeca3da064701bcf1b2586851b5935055. Both exact nodes failed locally on that base; their test files are unchanged by this PR.

Test Evidence and next action
test_app_mcp_scoping.py::TestAutoApproveIsFilteredAtTheWriteChokepoint::test_both_config_writers_run_the_pass The assertion searches _register_agents for the filter name, but commit 77aa76805ee4 moved filtering into render_app_agent_spec, which the writer calls before writing. The behavioral probe still strips governed autoApprove. Update the assertion to follow the actual call path; this does not justify duplicating or weakening the governance filter.
test_config_agent_fields_redaction.py::test_every_unguarded_str_field_is_in_the_masked_set The gallery commit adds source_app without adding it to the response-mask set. A further probe shows the loader also drops benign source_app provenance on read/save. The actual loader-to-GET probe did not expose a credential because the loader currently drops the value. Restore loading and add response masking together, so fixing provenance does not open that latent exposure.

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, test_memory_multistore_stress.py::test_private_store_mixed_load[32], has a separate fixture-order dependency inherited from merged #10586. The unmodified focused test passes on both revisions; a bounded diagnostic that lets recalls complete before global fact writes reproduces 32/34 on both revisions. Only the two Global V1 recalls miss; all private V2 recalls and the 255-operation/isolation/vector invariants pass. The episodic seed is not independently admissible to V1 retrieval. A fixture-only repair should seed an admissible record before concurrent work and retain the 34/34 assertion; changing runtime retrieval policy is not supported by this evidence. This reproduces the macOS signature, though the macOS log does not record its exact scheduling.

The fourth macOS failure was introduced here: a raw model-namespace literal violated the identity check. That comparison is repaired in e98203fe023fa34ce094d5038cc77096ec7f5192 using a named model-namespace constant. The final SDK/namespace run passes 336 tests, including both identity contracts. No platform-helper test failed in the completed macOS run.

@Zedmor

Zedmor commented Sep 16, 2026 •

Copy link
Copy Markdown
Contributor Author
  • macOS verification of process helpers span=3c79f410f5f6

Disposition: fixed by native verification of 68cdff832e4bb5fb5db700581a9643b1e221709f in https://github.com/kirodotdev/KiroCrew/actions/runs/35115715445. All three macOS shards completed their test suites: 106,183 passed; no platform-helper test failed. The platform test file is neither collect-ignored nor in the macOS expected-failure list.
The full run was not green: it exposed one namespace-literal violation repaired in e98203f and verified by both SDK identity contracts, two assertions reproduced on the unchanged lifecycle base, and a memory-stress fixture ordering failure reproduced on both revisions. Their evidence and dependency ownership are documented at #11324 (comment).
The platform implementation is unchanged by the review repairs. Linux additionally passes real parent-PID observation and a new regression that hides libc's renameat2 symbol, proves atomic publication and collision refusal, and refuses unsupported ABIs. That regression fails against the original lifecycle implementation.

@github-actions github-actions Bot added readiness: checking Automated validation is still running and removed readiness: passed Eligible automated validation passed for the current revision labels Sep 16, 2026
@Zedmor
Zedmor force-pushed the feat/member-model-backend branch from 68cdff8 to e98203f Compare September 16, 2026 17:16
@Zedmor

Zedmor commented Sep 16, 2026

Copy link
Copy Markdown
Contributor Author
  • Preserve pins while exploring harness choices span=3ed80eb3e2ae

Disposition: fixed in e98203f. The editor retains compatible drafts, restores the opening model and effort when returning to the original backend, and treats inherited/explicit choices resolving to the same backend as equivalent. It waits through errors and auto-only polling before declaring a pin incompatible; late responses cannot clear restored pins.
Regression coverage includes incompatible and compatible catalogs arriving after the real polling interval, equivalent routes, roster refreshes and save payloads. A real browser also exercised away/back and save/reopen on desktop and mobile.

@Zedmor

Zedmor commented Sep 16, 2026

Copy link
Copy Markdown
Contributor Author
  • Explain the backend selector span=2747803d4fb3

Disposition: fixed in e98203f. The selector now uses the existing translated Kiro CLI, Claude Code and KAS display names and lifecycle helper text explaining that an already-open session keeps its harness. Registered backends without a frontend translation retain their policy name.
Desktop and 390px mobile captures show the new labels and helper in context. This reuses existing catalog entries; static and rendered translation checks passed.

@Zedmor

Zedmor commented Sep 16, 2026

Copy link
Copy Markdown
Contributor Author
  • Localize catalog failures and retry in place span=6816670165e9

Disposition: fixed in e98203f. Backend and model fetch failures now render existing localized error strings beside the affected field, with an explicit Retry action. They do not display raw transport exception text or navigate away from the draft.
Tests cover both failing lists, retry success and draft preservation. The Retry button disables during its own fetch; the model query waits until its inherited route is known.

@Zedmor

Zedmor commented Sep 16, 2026

Copy link
Copy Markdown
Contributor Author
  • Validate private admission against the proposed inherited route span=583f63beb6ea

Disposition: fixed in e98203f. Private-memory admission now checks the proposed effective backend before a configuration write or provisioning. Clearing a pin respects the member-conversation backend for verified enrollment and the global route for plain workers; explicit Kiro remains an explicit selection. Model catalogs and validation use the same route distinction.
Tests cover both global/member backend directions, explicit Kiro, provisioning, private unenrolled workers, unreadable enrollment and foreign active-provider namespaces. Codex private memory is still refused; no authentication path or downgrade was added.

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Sep 16, 2026
@Zedmor

Zedmor commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author
  • Unsupported saved-member capability adoption. span=8dda91601b33

Disposition: Retained as a documented capability limitation, with earlier save-time feedback acknowledged as a useful advisory. An explicitly enrolled member can save an app that lacks full saved-spec support, and its next allocation then refuses with capability_harness_unsupported before starting the provider. That observation is correct. Ordinary member creation does not enroll capability inheritance; enrollment is an owner opt-in. Existing conversations retain their running app.

The existing contract separates saved configuration from verified runtime adoption: agent_capabilities.py requires explicit enrollment, crew-mode.md documents pending/unverified adoption and this unsupported-startup refusal, and H6 requires opt-in full-spec support. The real provider-support and zero-start tests exercise that refusal. No spec is silently dropped and no unsupported provider is admitted. This PR does not claim full-spec support for every selectable app.

A save-time check or selector explanation could surface this limitation sooner; it would be an additional acceptance/feedback policy, not a replacement for runtime admission. I am retaining the current documented contract for this revision. The independent source review found no mandatory correctness repair here; it did not execute a combined enrolled save-to-startup browser scenario, so this reply does not claim that additional test.

@Zedmor

Zedmor commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author
  • Shared namespace predicate suggestion. span=e42c562f6f0f

Disposition: Retained as a maintainability advisory. The foreign-member predicate is duplicated in the runtime and display resolvers; the recommendation can reduce future drift. No current divergence was demonstrated. The resolvers receive different identity/context inputs and preserve different output spellings, so any extraction should share only the predicate rather than collapse their resolution paths.

The contextual readout, same-namespace template and member-execution tests cover agreement on the behavior this PR changes. I am keeping the two small predicates in this revision; this is not a claim that duplication is inherently preferable or that an extraction is already scheduled.

@Zedmor

Zedmor commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author
  • Installation-default inheritance versus model compatibility. span=5d900309ee8f

Disposition: Retained with an explicit policy rationale. The rule answers which defaults an explicitly selected member app inherits, while scoped_pin answers whether a particular pin belongs to the selected namespace. The intended isolation is that a member explicitly moved to another model namespace uses its own pin or that app's Auto; an installation template/global/role pin should not replace that choice merely because both apps accept the string. Inherited app selection retains the installation fallback, and explicit per-call model choices still take precedence.

The synthetic resolution probe demonstrates the distinction: with an installation probe-global pin, explicit Codex plus no member pin resolves to Auto in the readout, chat and delegated paths; inherited Codex routing retains probe-global. This proves policy consistency, not native incompatibility or a measured cost difference. The related member-selection tests pin the same boundary.

The reviewer is right that a both-scoped model is not necessarily invalid on the other app. I am not using invalidity as the justification. Removing the branches changes the documented inheritance policy into a compatibility-only policy; that is a product choice for the existing human review, not a demonstrated resolver correction. No override or maintainer approval is asserted here.

@Zedmor

Zedmor commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author
  • Recovery and admission preserve an execution selection. span=6d3e0be1769f

Disposition: Retained with evidence limits. The concrete inconsistency is that main can carry a requested model from an existing run across queued-work reset while reselecting its backend from a newly changed default. The clean-base probe exercises that manager path and observes the old model forwarded to the new backend. This branch keeps the backend/model pair together for reset and failed-compaction successors; a genuinely fresh allocation observes current defaults. The same captured tuple keeps admitted work and continuation from re-reading only some of those choices later.

This applies to unpinned members too because the inconsistent pair occurs when the installation default changes; limiting the correction to an explicit member backend would retain that path. The regression distinguishes live reuse, queued reset, failed-compaction restart and a fresh allocation. It is a provider-double forwarding proof, not evidence that a particular native model was rejected. Native Kiro/Claude Auto tasks passed separately; native Codex inference remains unverified due to sign-in refusal.

The draft backend-selection RFC is not an accepted decision, and the accepted Profile/template RFCs support UI placement and member/template identity, not this recovery policy. I will narrow that citation wording in the test-only CI repair revision. This is the implementation rationale for human review, not a claim that a native failure issue or maintainer decision already exists; no review override is being used.

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.
@Zedmor

Zedmor commented Oct 11, 2026 •

Copy link
Copy Markdown
Contributor Author

CI repairs in 754eafcfca3f3ab3beb09d24fb5d428896096ca2:

  • AcpAdapter.defaults.test.ts now checks the optional routing context, including absent context, normal chat and literal templates. The former one-argument spy expectation reproduced the CI failure. The correction passes 660 executions across focused, 20 serial, 10 parallel and shuffled runs.
  • test_slot_running_readers.py explicitly declares _slot_selection_backend_guard as a reservation reader and adds it to the helper map so future callers remain enumerated. Both Linux and Windows failed on that single missing declaration. The scanner and exact-set assertion remain intact. The correction passes 92 executions on the same repeat/shuffle schedule beside busy/allocation-race regressions.

Production and UI files are unchanged from 6e0bedb842652. The periodic mechanism retrospective recommends retaining the current design; no additional production changes were introduced. Fresh CI run 38170059148 passed on this exact head: all Linux and Windows backend shards, frontend shards, integration/E2E checks and coverage gates are green. GPT and Opus reviews both passed on this exact head with no findings. The prior failed checks are superseded by this successful run, not treated as passes.

@Zedmor

Zedmor commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author
  • Shared inheritance predicate and cross-path coverage. span=59abc3aa3c5e

Disposition: Accepted as a maintainability and coverage advisory; no present divergence has been demonstrated. The two loader paths repeat the explicit-foreign-member eligibility predicate, and sharing that predicate could reduce drift. The other two sites are not equivalent complete resolvers: _session_model returns a caller-side input or None so the factory can finish resolution, while _capture_member_settings uses the actual child backend to select delegated role defaults and then calls acp_effective_model. Its effort precedence intentionally includes the subagent role before resolve_session_effort. A shared eligibility predicate is feasible; session routing, actual backend, role precedence and final output spelling must remain separate inputs to the callers. The proposed predicate extraction can preserve those semantics.

The readout/factory comparison is already exercised by test_model_readout_uses_the_session_route and the literal-template tests. Separate member-capture cases pin member precedence, explicit Auto/effort, and omission of template/global/role models for an explicit foreign-app member. The earlier resolution probe compares readout, factory and capture: explicit foreign app plus no member model gives Auto in all three; inherited routing with the same explicit child backend gives the installation fallback in all three. Without that override the child correctly follows the global app and subagent role while the DM follows member_acp_backend. These are resolution/factory proofs, not native inference.

There is no single exhaustive null/empty/other-backend by set/unset-model by model/effort matrix, so I am not claiming the requested coverage already exists. Such a table is a useful optional improvement if it fixes session context, actual backend and role inputs, normalizes display versus wire spellings, and explicitly preserves the expected role differences. I am retaining the current paths and tests for this test-only revision, without claiming duplication is preferable or extraction is scheduled.

@Zedmor

Zedmor commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author
  • Unpinned member restart captures its running selection. span=778238d51aab

Disposition: Retained with the evidence boundary explicit. The change applies to unpinned members, and that compatibility change is disclosed. The observed defect is a partially retained execution selection: after the installation member default changes, queued reset on clean 7bb9d57 forwards the original requested model while reselecting a different backend. The portable test uses the production factory, allocation, identity forwarding and queued reset, with provider construction/start doubled. It observes (Kiro, probe-kiro-model) becoming (Codex, probe-kiro-model); recording the started backend on that same isolated base session makes the control pass. Current allocation keeps the pair, while a genuinely fresh allocation follows the new default.

The committed eight-case test test_member_live_reuse_and_captured_restarts_keep_backend_after_defaults_change distinguishes profile/default edits, queued reset/failed-compaction re-entry and configured-model/Auto cases, and also checks fresh allocation. The member-default cases deliberately keep the member backend null. Restricting capture to non-null member fields at initial allocation removes the recorded backend precisely for those cases and restores partial re-resolution. The capture uses the actual started backend and factory model after pool selection; it does not freeze installation defaults for newly allocated conversations.

The reviewer is right about the remaining evidence limit: no linked end-user issue or native successor rejection is supplied, and the portable base comparison was 7bb9d57, not a newly rerun 64e1cbb base. The synthetic probe does not satisfy a demand for live model rejection; it establishes the forwarding inconsistency and its counterfactual repair. Keeping a running execution selection together is the rationale for retaining this behavior, not a claim of native failure, maintainer approval or review override. The sixth description has already narrowed the RFC citations: a draft backend-selection RFC is not an accepted decision, and the Profile/template RFCs are not evidence for this recovery policy.

@Zedmor

Zedmor commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author
  • Default versus explicit app wording. span=627ea7982f15

Disposition: Accepted as a wording advisory. The two rows make different persisted choices: Use default writes null and an app row writes that explicit backend. They can resolve to the same app today and differ for fresh allocations after the default changes; existing conversations retain their captured app. The current hint explains following the default versus pinning, but the blind-reader feedback shows that the labels and hint did not make this clear to that reader. The selection/round-trip tests establish different semantics, not comprehension. More explicit labels are a reasonable small improvement; no incorrect value or lost choice is demonstrated here. Retaining the current wording in this test-only revision is not a claim that the reader understood it.

@Zedmor

Zedmor commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author
  • Profile saved-status explanation. span=70f5ecb472d5

Disposition: Accepted as a discoverability/wording advisory. Profile has model and effort controls but no app control, so referring to the app requires context the panel does not supply. The status describes the actual save outcome: the member preference was saved, while an unknown or different slot backend prevents this Profile action from applying it to that conversation. Matching slots can receive the guarded update, and real request failures remain errors. The proposed model-specific replacement would not cover effort-only saves or clearing a model to inherit; any replacement should describe the saved preference and unchanged conversation generically. The existing status is truthful, though the reader feedback supports clearer wording. No mandatory runtime correction is identified.

@Zedmor

Zedmor commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author
  • Retry button visual affordance. span=710f14ae659c

Disposition: Accepted as a visual-affordance advisory. The flex-column wrapper stretches the shared, outlined Btn, and the reader mistook it for an input. It is an actual button whose handler refetches both the app list and backend-scoped model catalog without clearing the draft; the earlier controlled-failure QA exercised those calls and draft retention. That functional evidence does not rebut the reported visual ambiguity. Giving this instance self-start is a small optional presentation improvement; a new icon is not necessary to establish correct retry behavior. No nonfunctional retry or draft loss is demonstrated.

@Zedmor

Zedmor commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author
  • AI app label style. span=cf2a82530e5a

Disposition: Accepted as a new visual-consistency advisory. The AI app label uses text-[12px] text-muted, while the shared Field used by Model and Reasoning effort uses uppercase, tracking-wider and font-medium. The inconsistency is real. Matching those label classes would be a small cosmetic change; the existing htmlFor/id association should be preserved. This does not demonstrate a broken selection, missing accessible name or incorrect saved value, and no mandatory functional fix is identified for this test-only revision.

@Zedmor

Zedmor commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author
  • Plain-language app hint. span=b3a05abd26c0

Disposition: Accepted as an editorial advisory. The reader did not understand pins and installation, and member versus crewmate also deserves consistent wording. The proposed sentence is useful for timing but, by itself, omits the distinction between following a changing default and fixing an explicit app, which is the first concern in this review. A future wording choice should explain both selection semantics and the fresh-allocation boundary; open conversations retain their app. The current hint supplies both facts, but its completeness does not prove clarity. No wording work is asserted to be scheduled.

@Zedmor

Zedmor commented Oct 11, 2026

Copy link
Copy Markdown
Contributor Author
  • Saved readout beside an unsaved model draft. span=a19badd8eae5

Disposition: Accepted as a saved-versus-draft presentation advisory. The fields show the editable draft while the readout still reports the saved resolution; changing the backend field hides the readout because it describes the saved backend selection, even when the draft currently resolves to the same app. On model-only edits the two values can visibly differ, so the reader uncertainty is valid. Save/reopen tests prove persistence and do not prove that Using/Thinking at distinguishes saved state clearly enough. Labeling the readout Saved is a proportional optional remedy; hiding it on every draft edit is not required for correctness. No incorrect persistence or resolution is demonstrated by this observation.

@bolichen97

Copy link
Copy Markdown
Contributor

[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 agents.<name>.acp_backend key, so only one can land.
Why (a): it is the superset. Only this PR scopes the model and effort to the pinned harness and has the dashboard selector. The pin stays owner-only (_require_owner), outside members.SELF_EDITABLE_FIELDS, and goes through resolve_selected_backend, so it adds no agent power.

Next, before merge (rebase needed, 24 files conflict):

  1. Store the field in main's harness vocabulary from #20359: read through backend_from_stored, write through backend_to_stored; "kiro" pins Kiro CLI, null/"" inherit.
  2. Route every select_provider_backend( reader through one crew-aware helper, with a test that fails on a new call site (the one idea taken from feat(crew): let a crew member pin its own ACP backend #16200).
  3. Precedence stays per-call backend= > member pin > member DM route > default.
  4. Follow feat(crew): let a crew member pin its own ACP backend #16200's rule: do not return acp_backend from GET /api/agents.
  5. Give the schema field the same values_fn enum as agent.acp_backend.
    iamwhatever's request for review from Zezhen and Serena still stands. This ruling does not replace it.

@bolichen97

Copy link
Copy Markdown
Contributor

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:

  1. The changed-paths hot-reload rule and its config watch tests: when an agents.<name>.acp_backend edit is deferred, retried or folded in, the provider factory is rebuilt, so the pin edit is not lost.
  2. A call-site guard test that fails when any select_provider_backend( caller skips the member pin.
  3. Keep acp_backend out of the GET /api/agents response.
  4. Move the dashboard selector off CrewEditorDialog.tsx, which main has deleted, into the new crew editor.

#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.

@iamwhatever

Copy link
Copy Markdown
Contributor

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 loader.py into loader_config.py, CrewEditorDialog.tsx deleted for the new crew editor). This is a re-port, not a conflict fix. Before that work starts, two things need a maintainer:

  1. Read path for the editor (design decision). The ruling asks to keep acp_backend out of GET /api/agents (ask 3/4). But the new crew editor (useCrewEditor.ts) seeds every field from the roster record that GET /api/agents returns, and GET /api/config/kirocrew masks the field as agent-writable free text (per feat(crew): let a crew member pin its own ACP backend #16200). So the selector has no place to read the stored pin. Please pick one:
    • (a) return acp_backend from GET /api/agents after all, now that a dashboard control reads it (feat(crew): let a crew member pin its own ACP backend #16200 said it was withheld "until a dashboard control reads it");
    • (b) add it to the owner-only GET /api/agents/{name}/resolved-model readout the editor already calls;
    • (c) a new owner-only read endpoint.
  2. Human review. The CHANGES_REQUESTED review asking for review from Zezhen and Serena still stands, and the ruling says it is not replaced. Only they can clear it.

Once (1) is decided, the rest of the ruling (harness vocabulary from #20359, one crew-aware select_provider_backend helper with a call-site guard test, precedence, values_fn enum, hot-reload rule from #16200, selector in the new editor) is clear enough to implement.

@bolichen97 bolichen97 left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[operator: bolichen97] My earlier change request is addressed on 754eafc. Approving to clear it; other open gates on this PR still apply.

@kyleseaman

Copy link
Copy Markdown
Contributor

I'm re-porting this onto current main as a new PR, scoped to crewmates: a per-crewmate harness pin (agents.<name>.acp_backend) next to the existing model pin, shown only behind a developer feature preview. @Zedmor is credited as co-author on the commit and in the PR body. I'll link the new PR here once it's open.

@kyleseaman

Copy link
Copy Markdown
Contributor

The re-port is open: #25485. It pins a mate's harness (agents.<name>.acp_backend) beside its existing model pin, fails closed when the harness cannot start, and shows the control behind a Feature Previews switch. @Zedmor is co-author on the commit.

@CrysisDeu CrysisDeu mentioned this pull request Oct 29, 2026
5 tasks done

@CrysisDeu CrysisDeu left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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:

  1. Blank vocabulary follows main (#20359): store through backend_to_stored, read through backend_from_stored; ""/absent = no choice, "kiro" = Kiro. This PR's "" = explicit kiro-cli is out.
  2. 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's slot_backend (per-chat tier): precedence spawn > chat pick > member pin > member route > default.
  3. Editor read path: (b), the owner-gated resolved-model readout; acp_backend stays out of GET /api/agents.
  4. A crewmate's pin is validated at write time against ACP_BACKENDS_MATE_WORKER_KIT (today kas, claude), and crewmate_worker_backend reads the member pin. A pin outside the kit is refused with a message naming the setting, not agent.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 --agent off disk, no per-session channel). A member "kiro" pin keeps main's remap to KAS.
  5. Missing or denied pinned backend fails closed with a visible refusal for enrolled members; no silent degrade to Kiro.
  6. Carry over from #16200: the changed-paths hot-reload rule and its ConfigWatch tests, 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.

@bolichen97

Copy link
Copy Markdown
Contributor

@Zedmor this branch now conflicts with main in about 30 files. They include config/loader.py, config/sections.py, session.py, session_allocation.py, the subagent manager and the crew editor. Main has taken several large merges in those areas since your last push on 2026-10-11, so the conflicts touch real logic, such as acp_effective_model and the member backend resolution, not only moved lines.

The maintainer ruling still stands: this PR is the path for the per-member harness pin. Could you rebase it onto current main when you have a moment? A full review runs as soon as the new head is up.

@iamwhatever

Copy link
Copy Markdown
Contributor

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 (agents.<name>.acp_backend) next to the model pin, with harness-scoped models, a fail-closed check and the profile control. As @CrysisDeu noted on 2026-10-29, this PR closes once #25485 lands, so I am closing it now. Your branch is kept. #25485 is scoped to a crewmate's own chat and its worker, so the delegated-run capture and recovery-path parts of this PR were not carried over. If you still want any of that, or anything else is missing, please reopen this PR or open a small follow-up on top of main.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

merge conflict Branch has merge conflicts with its base — author must resolve before merge

Projects

None yet

Development

Successfully merging this pull request may close these issues.

5 participants