Repository navigation
fix(dashboard): continue after false tool blocker - #9748
NicholasRBowers merged 1 commit into
Conversation
|
Intent: Keep an attended dashboard request moving when the model falsely claims the current turn lost tool execution after read-only preparation, without granting authority from model-authored or fetched text. |
|
👋 Hi! This PR's description is missing some required sections from our PR template. Workflow runs won't be auto-approved until the description is updated. Missing sections:
Please update your PR description to include these sections, then push or re-save the description. The workflows will be approved on the next cycle. |
Design Review (Fable 5, fork) — 🟡 CONCERNSDesign-level review of Design-Verdict: CONCERNS Undiagnosed model failure patched with a two-sentence exact-match grammar and no linked diagnosis issue — the interim mitigation has no exit owner. Watch
[DESIGN-REVIEWED] ea1260f |
First Principles Review (Fable 5, fork) — 🟡 CONCERNSPremise-level review of First-Principles-Verdict: CONCERNS The defect and the introspect/tool_search read-only claim rest only on this PR's own description — no linked issue, transcript, or governance pin. Not justified as shipped
What this change shipsInventory (10 items) — 7 justifiedIntent: stop a Kiro turn that falsely claims tool loss after read-only preparation from forcing the user to type "continue" — a FIX.
[FIRST-PRINCIPLES-REVIEWED] ea1260f |
Opus 4.8 Review (fork) — ✅ no blocking findingsReviewed Review detailsI have the full applied picture now (working tree is at base
The candidate list contained no candidates to falsify, and I could not independently ground any behavioral defect or rule violation on the added/changed lines to the 80+ bar. No findings. [OPUS-REVIEWED] ea1260f |
GPT 5.6 Review (fork) — ✅ no blocking findings (all downgraded on adjudication)Reviewed Opus 4.8 adjudication downgraded every blocking finding to advisory. Adjudication downgraded all 1 blocking finding(s) to advisory: the remedy each one requires is disproportionate to the harm it prevents. Read them as advice, not as merge conditions. Review detailsBLOCKING -- src/kiro_crew/dashboard/chat_runner.py:9082 -- Tool evidence retains unbounded provider-controlled data
Adjudication (Opus 4.8) — is blocking on each finding proportionate?Per-turn locals in Harm rung: LOW. (a) Growth is bounded by the number of tool calls in ONE turn — these are function locals, not slot/session state ( [ADJUDICATION] ea1260f total=1 uphold=0 downgrade=1 |
ec6c78d to
0008f04
Compare
|
Addressed the prior design/first-principles concerns on the new revision:
Local mutation proof and both CI-contract review lanes are green on |
Commit |
|
Design concern disposition — reviewed head
These clarifications landed in |
|
First-Principles concern disposition — reviewed head
The reconciliation and scope wording landed in |
75b7d21 to
c06225c
Compare
|
self-added: yes
|
|
|
|
|
|
🤖 Kiro Crew [operator: NicholasRBowers#a942f9ca]: This PR was flagged as stale with a merge conflict and failing CI. I've assessed the blockers and they appear resolvable — I'll push fixes directly to this branch as a co-author. Assessment: GPT 5.6 and Opus 4.8 are both clean on this head, and every red traces to base drift rather than to this diff. (1) Two conflicts. In If you'd prefer I don't touch this PR, add the |
Drive status — REVIEW-READY ✅Head What this PR carries (single squashed commit, rubencu authorship + Kiro Crew co-author trailer):
Drive history: 2 upstream breakages ridden out (#12182 sandbox import, #12713 slot-reader enumeration — the latter unblocked via PR #12715 which prompted the maintainers' #12718), 3 installation rate-limit events cured by SHA re-mints, 6 Windows shard flakes, 1 E2E gateway-hang flake. All dispositions on the thread trace to verified evidence. (This comment is updated in place by the drive loop.) |
|
🤖 Kiro Crew [operator: NicholasRBowers#a942f9ca] Rebased: main moved in |
|
🤖 Kiro Crew [operator: NicholasRBowers#a942f9ca] Rebased: main moved +192 commits ( |
|
🤖 Kiro Crew [operator: NicholasRBowers#a942f9ca] Rebased: main moved |
|
|
|
|
|
|
|
|
|
|
Review fixes pushed: |
|
Rebased: |
|
Rebased: |
|
Re-minted (no code change): |
|
|
Security fix pushed: |
|
Lint fix + rebase: |
|
Rebased: |
Replay an authenticated user's request once when the Kiro backend stops after proven read-only preparation with a false current-turn tool blocker. Preserve attachment metadata and ledger actor provenance across the replay, and keep provider tool identity explicit and fail-closed.\n\nRefresh the Goose ACP event snapshots for the trusted identity field emitted by its canonical tool-call metadata. Rebased onto current main by Kiro Crew driving this PR to green: resolved merge conflicts in chat_runner.py (kept main's requeue-helper structure -- extra_meta param, _recovery_qid capture, refusal-retry block, _crew_log_actor rename -- with this PR's recovery_meta enrichment layered in) and session.md (kept both new subsections), and pruned test/test_acp_tool_identity.py from the black baseline now that this PR leaves it black-clean. Original work by rubencu. Review-round fixes: the false-blocker replay arm now snapshots the session-scoped stop counter at enqueue (same as its sibling recovery arms), so the dispatch-point purge sees a Stop issued on a linked channel surface while the replay waited; adds a regression test pinning the enqueue-time snapshot and the purge. Corrects the session.md eligibility sentence to match the status-only completion acceptance (a terminal status=completed result decides eligibility, not output presence). Co-authored-by: Kiro Crew <noreply@kirodotdev.github.io>
|
Re-minted (no code change): |
|
Problem / Motivation
A Kiro-backed turn can complete every proven read/search/fetch preparation call and then falsely claim that a deferred tool blocked it. The user must type
continueeven though no real tool blocker exists.Why it matters
The false blocker interrupts a valid turn and teaches users to retry manually. Automatic recovery must remain bounded so model-authored prose cannot create replay authority, repeat completed tools, or loop forever.
What changed (motivation → approach → change)
Kiro Crew recognizes only the two observed complete false-blocker grammars after every unique call ID has a completed result from the fixed read-only preparation allowlist. It requires adapter-authored canonical tool identity, authenticated-human provenance, and the existing Stop/follow-up/steer/stage/auto-approval gates before replaying the original request once. A second blocker consumes the one-shot budget and stops.
Completion proof follows the ACP event contract: a provider terminal
status=completedis successful completion evidence even when the update's optional output payload is absent or structurally empty. The gate therefore does not require non-emptytool_output; failed, non-terminal, missing-ID, duplicate-ID, or count-mismatched calls remain ineligible. Requiring result text would reject valid empty successful tool results without adding provenance beyond the provider status.The replay preserves validated file/directory attachments and ledger actor provenance. A dedicated queue kind lets later user intervention purge it without treating model blocker text or enriched app context as user authority.
The attachment-preservation hunk in
_queue_recoverykeys onpayload == RecoveryPayload.ORIGINAL, so it also covers the pre-existing empty-response replay: that path's requeued row now carries the triggering row's attachment lists too, where it previously fell back to whitespace-bounded attachment-name parsing. This is a deliberate shared-path change, not scoped to the new false-blocker replay.Blocker-adjacent wording that misses the two exact grammars never authorizes replay. After otherwise proven read-only preparation it emits an existing text-free warning containing only the completed-call count and slot identity, so wording drift is observable without logging model content or widening authority.
ACP events now separate adapter-authored canonical tool-name trust from complete MCP-pair trust.
tool_identity_trustedmay be true for a built-in name without an MCP server;mcp_identity_trustedstill requires a verified server/tool pair. The owningAcpEventdocumentation names both measured trusted harness channels,_meta.kiroand_meta.goose; no additional provenance is inferred.introspectandtool_searchremain explicit replay-only read/discovery entries rather than additions to the host auto-approval registry. Changing those entries or claiming broader upstream identity provenance remains a maintainer/upstream-evidence decision, not a local inference.Only the two Goose snapshots with adapter-authored canonical identity gain
tool_identity_trusted: true; no raw frame, input, kind, order, permission event, or non-Goose event changes.Implementation areas:
src/kiro_crew/acp/owns canonical identity extraction and event provenance;src/kiro_crew/dashboard/owns bounded replay and cancellation;src/kiro_crew/providers/propagates the trusted fields;docs/system-specs/modules/session.mdowns the contract;test/andtest/fixtures/acp_frames/own replay, identity, and byte-stability evidence..github/black-baseline.txtdropstest/test_acp_tool_identity.py, which this diff leaves black-clean, so the CI black gate's shrinking baseline stays accurate.Tests
ea1260fdfis one commit rebased onto basec293d24ae5ba97dd737d429d493239deb964b7af.git diff --check, and the single-commit push guard pass.Manual verification
N/A — unit and captured-frame coverage exercise the event contract and replay boundary directly. Broader adapter provenance and the replay-only inclusion of
introspect/tool_searchremain maintainer/upstream-evidence decisions; this candidate neither invents provenance nor removes either tool.Related Issues
no linked issue: bounded mitigation for an unsupported model self-report.
Pattern harvest
Rule candidate: review-prompt
Pattern: model-authored blocker recovery requires canonical tool identity, completed-result coverage, human provenance, and a one-shot budget.
Checklist