Repository navigation
feat(acp): run a Kiro CLI choice on KAS when kiro-cli offers v3 - #25350
Conversation
2c2a658 to
47f9e7b
Compare
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: |
First Principles Review (Opus 5.5) — ✅ human override accepted@buluoray overrode this lane for |
UX Review (Opus 5.5) — ✅ PASS (a BLOCK with no Trigger and no Outcome does not block)UX-level review of UX-Verdict: PASS cannot evaluate: missing the bell notification. The new bell note is how users first hear about the switch, and no screenshot shows it. The Agent Harness callout, button, saved state and row marker are all in the screenshots. Evidence gaps
Notes
[UX-REVIEWED] 56b4b1c Description read[DESCRIPTION-READ] a936d83f8e931700d1cac91debae9d0ff7f58a7b00596b95d3f2e4c28d783965 That sha256 names everything this verdict read: the PR description, plus the 6 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: the remap relies only on positive probe evidence, keeps the stored value, falls back to the unchanged v2 launch, and a revert undoes it cleanly. Notes
[DESIGN-REVIEWED] 56b4b1c Description read[DESCRIPTION-READ] a8955b460e9ad3cd3f73f9a42104e6f691f2eb95710641cb77352f89a9960934 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 |
Opus 5.5 Review — ✅ human override acceptedReviewed Human judgment by @buluoray overrides the Opus 5.5 finding for [OPUS-OVERRIDE] 56b4b1c Verdict recorded from an authorized human decision for commit False positive or not applicable? A repository writer can comment: |
47f9e7b to
c570674
Compare
c570674 to
ad6ce52
Compare
|
🤖 Kiro Crew Auto-Pipeline [operator: buluoray#ce473fca]: Owner decision, posted at @buluoray's direction. I confirm the product shape of this PR. kiro-cli makes v3 (KAS) its default engine on 11/2 and removes v2 at the end of November. So a user who picks Kiro CLI runs on the KAS path whenever their kiro-cli confirms it offers v3. Only an older kiro-cli that cannot offer v3 stays on v2. Crew never refuses to start a session over this. This is the intended change, not an accidental product shift. I am releasing the First Principles product-shape finding as the owner. |
ad6ce52 to
4cb2f47
Compare
233d3b5 to
db957f2
Compare
…er v3 kiro-cli switches a bare `kiro-cli acp` to its v3 (KAS) engine and is retiring v2. A stored `agent.acp_backend: "kiro"` now resolves to KAS at the one selection gate once the gateway's KAS floor confirmed, from a probe that ran, that the installed kiro-cli offers v3, and while KAS is selectable. The session then gets KAS's launch, projection, permissions and compaction. Without that evidence (another process, an unknown probe, a kiro-cli without v3, the bundle route, a policy that denies KAS) the pick keeps the v2 launch; no session is refused. - the gateway arms a Kiro CLI main backend on the floor and holds a stored pick until the verdict; a kiro-cli without v3 logs a notice naming `kiro-cli update` - one bell note, once per install, after a confirming verdict; the Agent Harness panel offers one click that saves `kas` - `kirocrew doctor` reports the KAS relay rows for a Kiro CLI pick and offers the kiro-cli update - decision entry docs/decisions/2026-10-28-a-kiro-cli-choice-runs-on-kas.md and the owning specs updated The Kiro CLI harness, skill-view aliases and cli.json overlay stay. Closes #25288
db957f2 to
56b4b1c
Compare
|
/ai-review override first-principles 56b4b1c owner-decision: The product shape is the owner's recorded choice: a Kiro CLI pick runs on KAS once kiro-cli confirms v3; old kiro-cli stays on v2; no session is refused. Owner decision: #25350 (comment) |
Human judgment recorded@buluoray recorded a owner-decision release (class
This decision applies only to this commit. A new push requires a new judgment. |
|
/ai-review override fable 56b4b1c owner-decision: Opus blocks only on governance, not code: the decision entry is signed by the project owner, and the Kiro pick now depends on the v3 probe. Both are the owner's recorded choice for the 11/2 KAS switch. Owner decision: #25350 (comment) |
Human judgment recorded@buluoray recorded a owner-decision release (class
This decision applies only to this commit. A new push requires a new judgment. |
Problem / Motivation
Goal: A user who picked the Kiro CLI harness keeps working when kiro-cli switches its default
acpengine to v3 (KAS) on 11/2 and removes v2 at the end of November.A Kiro CLI pick is launched as a bare
kiro-cli acp(src/kiro_crew/acp/client_spawn_env.py:87). Once kiro-cli's default flips, that process runs the v3 engine while Crew still drives it as the v2 harness: per-spawn skill-view aliases, the workspacecli.jsoneffort overlay and re-filtered agent specs, none of which v3 reads. Behaviour then depends on which side each feature consults.Why it matters
Every install that explicitly chose Kiro CLI (stored
agent.acp_backend: "kiro") is hit on 11/2, silently, and loses v2 entirely at the end of November.Not a goal
cli.jsonoverlay (follows once kiro-cli removes v2).What changed (motivation → approach → change)
Symptom → cause → change:
_read_backend_choice(config/loader_builders.py) reads a storedagent.acp_backend: "kiro"askiro_cli_pick_backend()(agent_sdk/backends.py): KAS once this process recordedconfirm_kiro_cli_offers_kas()and while KAS is selectable, Kiro CLI otherwise. The selection gateresolve_selected_backendis unchanged and answers Kiro CLI for Kiro CLI, so a Kiro CLI id it already answered (a degraded member floor, a worker backend) is never moved again. Every capability set reads KAS's membership for the moved pick, with no "picked Kiro" test anywhere (harness-parity H6/H7). Without the evidence — another process, an unknown probe, a kiro-cli without v3, the bundle route, a policy that denies KAS — the pick keeps the v2 launch. No session is refused.config.jsonkeeps"kiro".kas_engine_verdict()(agent_sdk/drivers/acp.py) splits theacp --helpprobe into supported / unsupported / unknown.apply_kas_cli_floorrecords the v3 confirmation on any "supported" probe (so a Kiro CLI pick made later in the process moves too) and re-resolves a stored pick to KAS; on "unsupported" it marks KAS unservable and logsKAS_CLI_FLOOR_KIRO_NOTICE(nameskiro-cli update); on "unknown" or the bundle route it changes nothing._start_kas_cli_floorarms a Kiro CLI main backend too; the off-loop declaration read releases it at once unless it is the stored pick, which waits (bounded by the probe timeout) so its first spawn lands on the verdict._reconcile_live_config_with_kiro_verdict) before it ends, so spawn waiters are released only afterSessionManagerrebuilt its provider factory from a fresh load (the diff namesagent.acp_backend), keeping any settings change made while the probe ran.get_or_createre-readsowner._provider_factoryafterawait_kas_floor_verdict, so a spawn that waited for the verdict uses the factory the verdict's reload installed. Not on the bundle route:_read_backend_choiceskips the remap whenagent.kas_bundleis on, and the notice and panel flag read the loaded field.spliceProbeRowkeeps the response's other fields.kiro_on_kas_noticewaits for the floor task, then posts one bell note (system.agent, links/settings/agent) only when the gate answers KAS for the stored pick; anO_EXCLmarker in the data home makes it once per install.GET /api/acp-backendsaddskiro_selection_runs_on_kas(same answer, so it is false until the floor confirmed v3), and the Agent Harness panel shows one line plus Save Kiro Agent as my setting (the option's visible name), which PATCHesagent.acp_backend=kasand leaves "Saved: Kiro Agent." in its place. The bell note uses the same name.validate_config_datareplaces a present non-stringagent.acp_backendwithMALFORMED_BACKEND_FLOORrather than"kiro", and_read_backend_choicereads it as the Kiro CLI floor, so a hand-edited bad value is never mistaken for a Kiro CLI pick and never moves onto KAS (harness-parity H3). The member field keeps its existing handling.AgentBackendTab.kiroOnKas.test.tsx"names both states and marks the Kiro CLI row", which reddens without the marker).kirocrew doctorprints the KAS relay rows for a stored Kiro CLI pick (not on the bundle route), and the engine row offerskiro-cli updateinstead of "switch to kiro".acp_backend_stored_as_kiro()(config/loader_document.py) is the one reader of the stored pick.knowledge/llm_pooldoes not warn when a Kiro CLI pick lands back on Kiro CLI there.docs/decisions/2026-10-28-a-kiro-cli-choice-runs-on-kas.md. Specs updated in the same commit: harness-parity H1, config, providers, acp-client, model-selection, agent-host-contract, slack-gateway; user docconfiguration.md.This changes what an existing setting does by default. There is no RFC on main for it; the owner's decision is issue #25288, recorded in the decision entry above.
Backwards compatibility
Compatible: no stored value is rewritten or rejected; a Kiro CLI pick still resolves to a startable harness everywhere (KAS only where the gateway confirmed v3, the unchanged Kiro CLI launch otherwise). In that gateway, readers of the resolved field see
kasfor the pick (the config response,to_dict()), the way a storedkiromember backend already reads as KAS.Removes nothing: the Kiro CLI harness, its launch, the skill-view aliases and the
cli.jsonoverlay all stay and still serve a Kiro CLI pick wherever v3 is not confirmed.Tests
Proof: not provable -- other: prove.py stops at BASELINE_RED because a changed test file carries host-only reds that also fail on main. Reverted by hand instead: removing the Kiro pick branch in
resolve_selected_backendreddenstest_acp_backend_kas.py::TestAKiroCliPickRunsOnKasOnceV3IsConfirmed::test_a_confirmed_v3_kiro_cli_runs_the_pick_on_kas,test_kas_cli_floor.py::TestDefaultNoticeWording::test_a_v3_kiro_cli_moves_the_pick_onto_kas,::TestAKiroCliPickWaitsForTheFloor::test_a_spawn_on_a_kiro_cli_pick_waits_and_lands_on_kasand the note tests; removing the panel callout reddensAgentBackendTab.kiroOnKas.test.tsx"offers one click that saves KAS as the setting".test_acp_backend_kas.py::TestAKiroCliPickRunsOnKasOnceV3IsConfirmed: no evidence → Kiro CLI; confirmed → KAS provider; policy denies KAS → Kiro CLI; a typo still takes the floor.test_acp_backend_kas.py::TestAKiroCliPickRunsOnKasOnceV3IsConfirmed::test_a_degraded_value_takes_the_floor_not_kascovers non-string values (7, true, [], null) with v3 confirmed; reverting the validation marker reddens 4 cases.::test_a_malformed_value_is_warned_about_oncepins a single warning.test_kas_cli_floor.py::TestAResolvedKiroIdIsNeverMovedAgain: a member floor from"not-a-backend"stays Kiro CLI through every re-resolution with v3 confirmed (reverting the gate change reddens 2).::TestTheVerdictReachesTheLiveFactory: a confirming verdict refreshes the live config before the floor task ends, no refresh without one, and the reload's diff namesagent.acp_backend(removing the refresh reddens 1).test_cli_doctor_agent_auth.py::TestDoctorKiroCliPickOnAnOldKiroCli::test_kas_only_rows_run_only_where_the_pick_runs_on_kas: a Kiro CLI pick the gateway keeps on v2 (unknown probe or no v3) gets no KAS permissions or tool-name doctor rows; a KAS pick and a pick on KAS still do (removing the skip reddens 2).test_kas_cli_floor.py::TestAWaitingColdStartUsesTheFactoryInForce: a factory swapped during the wait builds the provider (removing the re-read reddens it). Bundle-route tests intest_acp_backend_kas.pyandtest_kiro_on_kas_notice.py, and the vitest "keeps the callout when another harness is checked again", each redden without their fix. Test waits are bounded withasyncio.wait_for.test_kas_cli_floor.py: v3 moves the pick; an unknown probe or one without v3 does not; the bundle route never probes; a floor answer that is not a pick is not moved; the old-kiro-cli notice nameskiro-cli update(and the KAS member's notice too when both apply); gating per role; a spawn on the pick waits and lands on KAS.test_kiro_on_kas_notice.py: claimed once; nothing without a positive verdict, on an old kiro-cli, under a policy that denies KAS, or after a failed probe; waits for the floor; the panel flag reads the same answer.test_cli_doctor_agent_auth.py: the engine row offerskiro-cli updatefor a Kiro CLI pick; no KAS block for one on the bundle route.AgentBackendTab.kiroOnKas.test.tsx: the callout and its PATCH; nothing shown when the flag is false or absent.test/conftest.py: the floor's verdicts are per-test process state.Locally: 294 related backend test files (15,067 passed); every red there also fails on main on this host (Copilot
~/.mcp.json, systemd unit, guest Python). Related frontend tests,tsc -p tsconfig.app.json, black/isort/flake8/mypy, the harness-parity and comment-history gates clean.Manual verification
apply_kas_cli_floorone and each verdict is covered above.Screenshots / video
After pressing Save Kiro Agent as my setting
The bell note, open in the bell panel before the click (the real
NotificationSheetwith the note as the gateway posts it; title "Kiro CLI now runs on Kiro Agent", body "Kiro CLI now runs on Kiro Agent, so new sessions already use it. Open Agent Harness to save Kiro Agent as your setting.", link/settings/agent):Related Issues
Closes #25288
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)Contribution License Agreement
Contributions are accepted under the project's Apache-2.0 license.