Repository navigation
feat(crew): a Crew bot's harness follows Settings -> Agent Harness when unset - #29509
Conversation
UX Review (Opus 5.5) — ✅ PASSUX-level review of UX-Verdict: PASS Nothing blocks: screenshots 01–08 cover all eight new lines, the blind reader read each one correctly, and no element is swapped out. Notes
[UX-REVIEWED] 83cab22 Description read[DESCRIPTION-READ] 77e42d96a4179b33864c2041a5e1e0f451d801a4220a565cc9648ea19103d9b2 That sha256 names everything this verdict read: the PR description, plus the 11 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 |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: unset Bot harnesses now resolve once in the loader, every existing reader picks that up, and the fallback and pin behaviour match H17 and the RFC. Notes
[DESIGN-REVIEWED] 83cab22 Description read[DESCRIPTION-READ] c00444effe8709d4f80c85a9c890eac3c67baebb8c7f76d2e0bac05323e01860 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 |
First Principles Review (Opus 5.5) — ✅ human override accepted@iamwhatever overrode this lane for |
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: |
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: |
8c5e79d to
711be6d
Compare
711be6d to
5cb0906
Compare
5cb0906 to
485e542
Compare
485e542 to
c57901c
Compare
|
@buluoray rebase only: the head is now
No other code changed. Local GPT and Opus found nothing blocking, and the related tests pass (1519). Could you re-review? |
|
@buluoray this push is a rebase only. The head is now
No other code changed. The security package line-budget test now passes on this head, because #32252 is on main. Local GPT and Opus found nothing blocking, and the related tests pass (1610). Could you re-review? |
|
/ai-review override first-principles c8e5646 owner-decision: same owner decision as #29509 (comment) (RFC #29442 and its revision #30116 are merged; only the RFC status line still reads draft). |
Human judgment recorded@iamwhatever recorded a owner-decision release (class
This decision applies only to this commit. A new push requires a new judgment. |
buluoray
left a comment
There was a problem hiding this comment.
🤖 Kiro Crew Auto-Pipeline [operator: buluoray]:
APPROVE on c8e5646. Every item from the change request on 07285d9 is fixed at this head, and I found no new defect.
What it does for users: a Bot with no harness of its own now follows Settings → Agent Harness. A Claude Code or Codex user's Bots run there instead of staying on Kiro Agent. The Settings page says which harness Bots run on and why. The person whose Bots moved gets one bell note at the first Bot start. A Bot pinned to a harness Bots cannot fully run on now runs on Kiro Agent and shows an error saying so.
Earlier change request, item by item (checked at this head, not taken from the author's summary):
- Pins misdescribed: fixed. Backwards compatibility now has a "Behaviour change for existing pins" entry. It says pins to pi, deepseek, goose, opencode, strands or copilot now run on Kiro Agent with an error bell, that a new save rejects pi and deepseek, and that an unstartable pin still refuses. Source agrees.
CrewBackendPin.checked(crewmate_harness.py:166) sends a pin thatbots_cannot_run_onmatches to KAS through the gate and setsharness_fallback.resolvestill raisesCrewmateHarnessUnavailablefor a pin that cannot start (H17).CREWMATE_CAPABLE_BACKENDS(:697) addsACP_BACKENDS_MATE_WORKER_KIT, soharness_pin_rejected(:771) refuses pi and deepseek. "Removes nothing" and "a pin still wins" are gone. The PUT route only judgesacp_backendwhen the body carries it (crew_update.py:361), and the profile page sends one key per save. So a Bot with an old pi pin can still save its other settings. - Reader lines: added, under "Readers of what tightened". They cover the save check (crew update route →
harness_pin_rejected,test_a_harness_without_the_worker_kit_is_not_offered_nor_saved) and the pin fallback (crewmate_backend_pin→CrewBackendPin.checked,TestUnrunnablePin). I found no other writer of a Bot'sacp_backendpin in src: the crew update route is the only save path. - Release note: fixed. The second line now covers the pinned-Bot move and the route back.
- Coverage gap: closed.
TestColdStartRefresh::test_a_plain_bot_cold_start_sends_one_moved_notedrivesrefresh_session_verdicttwice and asserts one note.
Since the earlier head: the push is a rebase onto main, plus that one new test and the lazy-import count in test_dashboard_server_composition_contract.py (now 136). I compared the PR's own patch at both heads with range-diff, and nothing else changed. One question I checked: does an inherited member_acp_backend reach disk on a whole-config save and freeze the inheritance? It does not. save() restores the harness fields from the stored document (_restore_stored_harness_fields, loader_config.py), and the PATCH writers write deltas.
Evidence:
- Mutation 1: replacing the
deliver_movedcall inrefresh_session_verdict(crewmate_harness.py:636) with a no-op reddensTestColdStartRefresh::test_a_plain_bot_cold_start_sends_one_moved_note(assert 0 == 1). The other 111 tests in the file pass. Restored, tree clean. - Mutation 2: dropping the worker-kit term from
CREWMATE_CAPABLE_BACKENDSreddenstest_a_harness_without_the_worker_kit_is_not_offered_nor_saved[pi]and[deepseek]. Restored, tree clean. - Tests on the clean head: test_crewmate_harness_inherit.py, test_crewmate_harness.py, test_kas_cli_floor.py, test_mate_spec_policy.py, test_dashboard_server_composition_contract.py and test_config_live_server_appliers.py ran 470 passed, 0 failed. I did not run the full suite or the frontend tests this round.
- CI on this head: 85 success, 11 skipped, none failed. PR Readiness is success. The merge with main is clean (merge-tree exit 0). take_away_check.py passes locally against the current body: no take-away shape in the diff.
- Lanes: Opus and GPT found nothing blocking. Design and UX gave PASS. First Principles was released by the author as an owner decision. That release points to RFC #29442 and amendment #30116, which maintainers merged.
- Source-only: the Settings panel copy and the bell note text. I read them but did not render them.
Non-blocking follow-ups:
- The reader list is prose bullets. take-away-changes.md asks for
Reader: <path>:<symbol> -- <entry> -- <test>lines. Each bullet already has the path, the symbol and the test. Rewriting them in that form, with the entry named (crew page for the save check, chat and the spawn entry for the pin fallback), would match the rule. - The release note says the picker offers "only Kiro Agent, Claude Code and Codex". Where the host's kiro-cli is confirmed to serve v3, a Kiro CLI pin resolves to KAS, so
crewmate_harnesses()also lists Kiro CLI. That was already true on main, and the pin runs on Kiro Agent, so the note is close enough. A parenthetical would make it exact. - The UX lane's evidence gaps still stand. No screenshot shows the Bot profile's Inherited hint. The Agent Harness screenshots show the older "Mates" wording, not the "Bots" copy this commit ships.
- The Design lane notes that a pin to goose and similar harnesses now runs on Kiro Agent instead of refusing. That gives a second "pin not honoured" path next to H17. RFC amendment #30116 records this as the chosen behaviour, so I treat it as decided.
|
requesting name change and simplify harness selection setting UI. too wordy |
bolichen97
left a comment
There was a problem hiding this comment.
requesting name change and simplify harness selection setting UI. too wordy
|
@bolichen97 thanks for the review.
Re-requesting review from you and @buluoray. |
|
/ai-review override first-principles 78feaa1 owner-decision: same owner decision as #29509 (comment) (RFC #29442 and its revision #30116 are merged; only the RFC status line still reads draft). |
Human judgment recorded@iamwhatever recorded a owner-decision release (class
This decision applies only to this commit. A new push requires a new judgment. |
|
@bolichen97 the simpler Agent Harness setting UI is in #32849 (#32849). This PR now only does the rename to Crew bot. |
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
This head needs fresh UI evidence before review can finish.
-
website/src/pages/members/CrewProfileSettings.tsx:62andwebsite/src/i18n/locales/en.json:8388. The UX lane on 78feaa1 could not evaluate this head, and a verdict blocked by missing evidence counts as a BLOCK. Two gaps:- Shots 01, 02 and 04 show the old copy ("Mates run on Claude Code ... A mate can pin its own in its settings."). This head ships "Bots run on {{harness}}, the harness picked here. A Bot can pin its own in its settings."
- The new Inherited hint in Bot Profile > Settings > Agent harness has no screenshot. That is
inheritedHarnessHint, with the four source strings at en.json:9304-9307. A unit test does not show how it renders.
Smallest fix: retake shots 01, 02 and 04 on the current head. Then add one shot of the Bot Profile harness picker, open, with the Inherited hint for each source (settings, legacy, fallback, worker_spec). After that, re-run the UX lane. No code change is needed.
Optional:
src/kiro_crew/dashboard/handlers/agents.py:1751: the_pin_entitlement_backenddocstring still describes the namespace comparison this PR removed. The body now always resolves the Bot's own member route.src/kiro_crew/agent_sdk/backends_members.py:526: an inherited pi or deepseek now falls back to KAS. RFC 5.3 lists that option as "Not chosen", and amendment 5.8 leaves it to the owner. A link to that decision in the body would help.- Please name the merge order with #31801 and #32849, and say which PR decides how a
kiropin behaves under O7.
Owner decision on bolichen97's review: log and warning messages, docstrings, comments and spec prose that #29509 adds say 'Crew bot', not 'crewmate'. Code identifiers, module names, config keys and RFC slugs keep their names (main already uses them). English copy this PR adds writes 'bot' in lowercase after another word, per docs/decisions/2026-11-08-bot-is-lowercase-after-another-word.md (en.json, en-XA.json, bell notes, field help, log lines). The _pin_entitlement_backend docstring now describes what the body does: with no pin it always returns the Bot's own member route.
|
@KiroCrewCaptain thanks. Review 5816581816 is addressed on
This head also follows Re-requesting review from @KiroCrewCaptain, @bolichen97 and @buluoray. |
|
/ai-review override first-principles 83cab22 owner-decision: same owner decision as #29509 (comment) (RFC #29442 and its revision #30116 are merged; only the RFC status line still reads draft). |
Human judgment recorded@iamwhatever recorded a owner-decision release (class
This decision applies only to this commit. A new push requires a new judgment. |
|
This changes behaviour users already rely on, so it needs a design sign-off from a maintainer before it can merge. A maintainer will post the decision here. Review result on 83cab22: no blocking defect found. Upgrade behaviour changes for unpinned Bots and for existing pins. The RFC this PR implements is still marked draft on main, and a change request asking for simpler, shorter harness settings copy is still open. Earlier change request from this account, item by item:
Optional:
|
bolichen97
left a comment
There was a problem hiding this comment.
Approving head 83cab22, and recording the maintainer design sign-off this PR was waiting on.
- Name: done. The text this PR adds says "Crew bot", and the UI copy keeps "Bot" per the recorded decision.
- Simpler harness setting UI: I accept the split. That work lands in #32849, which rebases on this PR as the Merge order section says, so this PR is not held for it.
- Design: I accept both upgrade changes the description lists. On an install with no stored
agent.member_acp_backendwhose Settings harness is Claude Code or Codex, unpinned Bots move off KAS onto that harness, with the one-time moved note. A Bot pinned to pi, deepseek, goose, opencode, strands or copilot now runs on Kiro Agent with an error bell note naming the pin, and a new save rejects pi and deepseek. A storedagent.member_acp_backendand pins to Kiro Agent, Claude Code or Codex resolve as before.
Captain's optional notes (the bots_may_run_on_kas notice wording, the RFC 5.3/5.4 amendment for pi/deepseek, and the _crewmate_backend test that copies its body) are fine as follow-ups.
Problem / Motivation
Goal: A Crew bot with no harness of its own runs on the harness picked in Settings -> Agent Harness, and the person can see which harness Crew bots run on and why.
Today a Crew bot with no pin runs on
agent.member_acp_backend, whose default iskas. Picking Claude Code or Codex in Settings -> Agent Harness moves every chat except Crew bots, and nothing on the page says Crew bots stayed behind.Why it matters
Someone who picks a harness expects their Crew bots to use it. Today they silently keep running on KAS, so model lists, costs and tool behaviour differ between a Crew bot's chat and every other chat, with no hint why.
Not a goal
agent_sdk/backends_members.py, unchanged.What changed (motivation → approach → change)
Design:
docs/request-for-change/rfc-crewmate-harness-follows-settings.md(#29442).Resolution for a Crew bot's own chat and the workers it opens: Crew bot pin
agents.<name>.acp_backend> an explicitly storedagent.member_acp_backend(legacy) >agent.acp_backend. Kiro CLI reads as KAS, as a storedmember_acp_backend: "kiro"already does._read_backend_choice(config/loader_builders.py) now resolves an unsetmember_acp_backendfrom the already resolvedagent.acp_backendthroughcrewmate_inherited_backend(agent_sdk/backends_members.py). The loaded field stays the one answer, so its 20+ readers follow Settings with no edit and cannot disagree. The selection gate keeps its shape (H13).ACP_BACKENDS_MEMBER_CAPABILITIES∩ACP_BACKENDS_MEMBER_DISPATCH∩ACP_BACKENDS_MATE_WORKER_KITruns Crew bots on KAS. Today that is Claude Code, Codex and KAS. The fallback is one warning per harness per process (_note_member_fallback, H3) plus one bell note (src/kiro_crew/crewmate_harness_notice.py), sent at start (src/kiro_crew/slack/gateway_runtime/service_boot.py) and on config change (config_watch.py).ACP_BACKENDS_MATE_WORKER_TOOL_LIMITS) and itshooks/permissions/toolsSettings(applies_spec_policy), and the chat tool limits of its saved Capabilities spec (ACP_BACKENDS_MEMBER_TOOL_LIMITS). Such a Crew bot runs on KAS, chat and workers, with a log line, a bell note opening its page and readout sourceworker_spec(worker_spec_forces_kasinsrc/kiro_crew/crewmate_harness.py, fields fromkas_only_worker_fieldsinsrc/kiro_crew/mate_worker.py). The answer is held in memory, read off the loop at start and on config change; the event loop only reads it and refreshes stale entries in a thread.watch_documentinsrc/kiro_crew/config/live.pysees a stored-to-inherited change. Log text insrc/kiro_crew/members_parts/member_sessions.py.CrewBackendPin.checked); a pin that cannot start still refuses (H17). The Bot's model follows the harness it lands on, through the gate's backend.CREWMATE_CAPABLE_BACKENDSinsrc/kiro_crew/crewmate_harness.pynow also needsACP_BACKENDS_MATE_WORKER_KIT, socrewmate_harnesses()offers only harnesses a Bot can fully run (KAS, Claude Code, Codex). pi and deepseek leave the picker and a save rejects them. A hand-edited pin to them still runs on KAS with an error, as above. The picker and the pin fallback now read one rule, so no offered pick lands on KAS.crewmate_harness_notice.deliver_moved, fromrefresh_session_verdict), claimed with anO_EXCLmarker under the data home likekiro_on_kas_notice. Every bell note now names the route: Agent harness in the Bot's Profile → Settings on the Bots page.src/kiro_crew/session_allocation_parts/get_or_create.pyawaitsrefresh_session_verdictoff the loop before the factory; a template apply (src/kiro_crew/mate_worker_kit.py) is judged on the spec being written.kiro_spelling_applies(backend)judges on the harness the pin runs on, inherited included;_pin_entitlement_backendjudges an unpinned Bot on its own route, never the install's (src/kiro_crew/dashboard/handlers/agents.py,src/kiro_crew/dashboard/agent_admin/crew_update.py,src/kiro_crew/cli_commands_agent.py).inherited_harness_sourceinsrc/kiro_crew/crewmate_harness.pylabels the answerlegacy/settings/fallback, andcrewmate_inherited_sourceaddsworker_specper Crew bot (member_acp_backend_declared()tells a stored value from none).GET /api/acp-backendsreturns it ascrewmate_harness; the crew readout (crew_records.py) addsinherited_sourceandsettings_harness.docs/decisions/2026-10-30-the-standing-agent-is-called-a-bot.md. The Agent Harness panel gets one line under the picker naming the harness Crew bots run on and the setting that decided it; a fallback is a warning-colouredrole="alert". A Crew bot's Inherited option names the same source. Types inwebsite/src/api/client.ts,website/src/api/client/agents.ts,website/src/api/client/config.ts; copy inwebsite/src/i18n/locales/(bn.json...zh-CN.json).src/kiro_crew/connections/status.pyno longer says Crew bots always run on KAS. Field help,config-baseline.jsonand the config / crew-mode specs describe the new default.Backwards compatibility
Behaviour change for existing installs. An install with no stored
agent.member_acp_backendwhose Settings harness is Claude Code or Codex moves its unpinned Bots off KAS on upgrade, onto that harness (unless a Bot's worker spec needs KAS). Their model, tool gate and native conversation move with them; the next turn starts a new native session, and Crew's own transcript and memory carry over. Such a person gets one bell note, once per install, at the first Bot start, saying so and how to set a Bot back to KAS (Profile → Settings → Agent harness).Behaviour change for existing pins. A Bot pinned to pi, deepseek, goose, opencode, strands or copilot used to run on that harness. It now runs on Kiro Agent, with an error bell note that names the pin and the route to change it (
CrewBackendPin.checked). A new save rejects pi and deepseek, which the #29498 picker offered and saved (harness_pin_rejected, viaCREWMATE_CAPABLE_BACKENDSadding the worker-kit term; RFC O6/O7). A pin that cannot start here still refuses (H17).Unchanged: a stored
agent.member_acp_backend(includingkiro) resolves exactly as before. A pin to Kiro Agent, Claude Code or Codex still wins. On every other Settings harness, unpinned Bots run on KAS as before, now with a warning when that harness is not Kiro CLI or KAS. No key, field or route is removed, and the API responses only gain fields.Readers of what tightened:
PUT /api/agents/{name}(dashboard/agent_admin/crew_update.py→crewmate_harness.harness_pin_rejected) now refuses pi and deepseek. Covered bytest_crewmate_harness.py::TestCrewmateHarnesses::test_a_harness_without_the_worker_kit_is_not_offered_nor_saved.crewmate_backend_pin→CrewBackendPin.checked, which lands a pin to a harness Bots cannot run on at Kiro Agent. Covered bytest_crewmate_harness_inherit.py::TestUnrunnablePin.Release note
Tests
test/test_crewmate_harness_inherit.py:TestColdStartRefresh::test_a_plain_bot_cold_start_sends_one_moved_note(a Bot cold start throughrefresh_session_verdictrings exactly one moved note; fails whendeliver_movedis a no-op);test_moved_notice_shown_once_when_inherited_differs_from_kas; no moved note for a KAS Settings harness, a fallback, a pin or the legacy setting;test_no_bell_when_crewmates_follow_the_settings_harnessstill rings no fallback bell, and now asserts the moved note is owed.test/test_crewmate_harness.py: pi and deepseek are neither offered nor saved; no offered harness is inkas_pin_harnesses().test/test_crewmate_harness_inherit.py: each tier (settings, explicit legacy incl. a legacy value outside the inheritable set,""as unset), Kiro CLI -> KAS (also on the bundle route), fallback for goose/opencode/strands/copilot and for pi/deepseek (worker kit), with one warning line; an inherited harness never refuses a Crew bot's worker; policy-denied Settings harness and denied KAS landing on the floor, pin wins / unstartable pin still raisesCrewmateHarnessUnavailable, worker backend agrees with the readout, bell note once per harness / not for settings or legacy / again after moving away / still owed after a failed delivery, never raises, panel readout.test/test_crewmate_harness_inherit.py(per Crew bot): tool limits, hooks, permissions or toolsSettings each run the Crew bot on KAS through the gate,pinned_backendand the worker route; a plain Crew bot follows Settings; legacy setting and pin never moved; inherited KAS reads no file; no fallback when KAS is not selectable; logged once, never for a pinned readout; the event loop never reads the file and a background refresh fills memory; the startup refresh reads each Crew bot once; chat tool limits; the watch sees an inheritance change; readoutworker_spec; model judged on KAS;kas_only_worker_fieldsnames each field; spelling follows the effective backend; loader mark never written back; per-mate bell note.website/src/test/AgentBackendTab.crewmateHarness.test.tsx: the three lines and the alert role; nothing on an older gateway.website/src/pages/members/CrewProfileSettings.test.tsx: the Inherited hint names each source.Manual verification
Rendered both surfaces in a real browser with
/api/answered from fixtures: Settings -> Agent Harness on a Vite dev build, and Bot Profile -> Settings -> Agent harness on the built SPA. Each shot below is from this head.Screenshots / video
Bots follow the Settings harness (Claude Code):
Fallback: goose picked, Bots run on Kiro Agent, shown as a warning:
Legacy
agent.member_acp_backenddecides:Bot Profile -> Settings -> Agent harness, open, with the Inherited hint for each source:
settings, Inherited is the Settings harness:legacy, Inherited is the storedagent.member_acp_backend:fallback, the Settings harness is one Bots can't run on:worker_spec, this Bot's worker settings need KAS:Owner decision: inherited pi or deepseek runs on KAS
RFC 5.8 leaves an inherited pi or deepseek to the owner. The owner decided that a harness Bots cannot run on falls back to KAS with an alarm: owner decision, recorded as O2 and O7 in RFC #29442 and amendment #30116. pi and deepseek have no worker kit, so an inherited one takes that fallback (
crewmate_inherited_backendinsrc/kiro_crew/agent_sdk/backends_members.py), with the same warning as goose.Merge order
#31801 decides how a
kiropin behaves under O7: it makes a Botkiropin start kiro-cli and leaves "should it run on KAS" to the maintainer. This PR does not change akiropin (bots_cannot_run_onskips Kiro CLI, andCrewBackendPin.resolvekeeps its launch rule). It only maps an inherited Kiro CLI Settings harness to KAS (RFC 5.1).Related Issues
Part of #21910
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)