Repository navigation
Description
Classification
- Type: RESEARCH / architecture reference / reuse-first
- Priority: P2 (after core 007 run-evidence contracts)
- Status: NOT ADOPTED / KILL-FIRST / NO IMPLEMENTATION AUTHORIZED
- Source: kirodotdev/KiroCrew
- Research Page: KiroCrew — persistent agents, Memory V2, Research Lab и verifier gates
- Related: Run evidence v2: action-level records linked to findings, costs and verifier outcomes #42 (action evidence); Run semantics: replayable event reducer and fail-closed required-gate verdicts #43 (run-event reducer); Reference architecture review: extract transparent agent-loop patterns from waku-agent #57 (transparent waku-agent loop). Also OwnAudit #43 for fix proof packets.
Objective
Review specific reusable contracts from KiroCrew, not a proposal to replace the current agent harness. Source review (18–21 Nov 2026) confirms repository docs, specs, GitHub workflow structure, not an independently executed system. Separate documented mechanisms from tested behavioral guarantees.
KiroCrew combines persistent ACP sessions, scheduled work, checkpoints, Memory V2 and opt-in apps. Most valuable candidates for 007 are:
- Durable task transitions and recovery: identity of task/attempt/action, checkpoint consistency, resume after gateway interruption, explicit cancelled/blocked states.
- External deterministic verifier gates: Auto-Improvement uses calibrated measurement + deterministic keep/revert + draft PR rather than allowing proposing model to certify its own patch. Source: auto-improvement spec.
- Two-stage independent PR review: high-recall discovery → independent grounded validation; reviewer prompt read from base commit, no PR-side prompt rewriting. Source: claude-review.yml.
- Agent memory as proposed derived context: source/provenance, correction history, repo scopes, lesson origin, immutable facts versus inferred suggestions; memory must not become policy. Source: Memory V2 docs.
- Cheap monitoring without unnecessary LLM turns: poll PR/check status deterministically; wake model only when actionable change exists, with time/token/error budgets. Source: PR monitoring.
- Fix-loop/SZZ analytics with LLM narrative separate from numbers: deterministic metrics cannot be overwritten by summarization; source: fix-loop-analysis.yml.
Strict ownership / avoid overlapping work
- Run evidence v2: action-level records linked to findings, costs and verifier outcomes #42 owns action identities, idempotent side-effect evidence; do not create a parallel action schema.
- Run semantics: replayable event reducer and fail-closed required-gate verdicts #43 owns persisted run events and replay reducer; this issue only examines mechanisms and gaps.
- Reference architecture review: extract transparent agent-loop patterns from waku-agent #57 owns the basic provider-neutral agent-loop contract. Do not duplicate its observer/typed tool outcome design.
- OwnAudit #43 owns remediation proof semantics, not 007.
- AI-Assistent normative/approved knowledge and Demand Radar market evidence are not 007's persistent memory.
- Nothing authorizes a production integration, new framework, autonomous edits, or new memory-to-policy path.
Kill-first experiment (proposal, not a command to run)
Stage A: desk audit of current 007 capabilities against the source mechanisms and explicit ADOPT / ADAPT / REJECT / EXISTING / UNKNOWN matrix. For every claimed KiroCrew guarantee, identify implementation + test, or keep UNKNOWN. Record commit hashes and note ACP harness parity gaps (Nov 13 matrix).
Stage B, only if Stage A shows a concrete gap: isolated synthetic repository with deterministic tool, fixed test suite, no production credentials:
- Baseline 007 multi-step task with independent verifier, checkpoint, forced interruption and resume.
- Duplicate delivery of one side-effect invocation ID; ensure no duplicated mutation.
- Failed test / ambiguous tool completion / unknown model token accounting; verify failure remains visible.
- Optional comparable KiroCrew run only when permissions, model/provider, data and cost conditions are equivalent. Otherwise mark NON-COMPARABLE.
Record: task/attempt/action IDs, log/artifact references, replay proof, duplicate side effects, human approvals, failed gates, total wall-clock, model usage (known/unknown), overhead. Fail-closed vs resume-from-stale-state is a first-class outcome.
Acceptance
- Source-pinned capability/evidence matrix: doc claim, code symbol, test, measured/UNKNOWN.
- A bounded comparison against existing Run evidence v2: action-level records linked to findings, costs and verifier outcomes #42/Run semantics: replayable event reducer and fail-closed required-gate verdicts #43/Reference architecture review: extract transparent agent-loop patterns from waku-agent #57 design (no duplicate subsystem).
- Clear candidate invariants for restart/resume, typed failure and idempotent tool effects.
- Distinct independent-verifier and model-review roles; negative controls for false positives and prompt injection.
- GO / ADAPT / PARK / KILL based on measurable net-new value, not feature count.
- If all meaningful mechanics already exist, explicitly close as NO-CHANGE.
Non-goals
No autonomous self-modification, no release/merge, no unreviewed tools on source repos, no storing secrets/production transcripts, no claim that a Python-focused Auto-Improvement app directly works on Go/C#.
Activity
PhysShell commentedon Nov 21, 2026
Addendum / Clare Liguori: Kiro unified harness + reactive agents (21.11.2026)
Source-backed reference: research Page, Kiro unified harness engineering article, MCP Events incubation repo, talk archive.
Net-new distinction vs KiroCrew: Kiro unified its separate IDE (TS), CLI (Rust), and web (Python) harnesses, because permissions/session/compaction semantics diverged; it uses a capability-oriented Cedar policy instead of individual per-tool command matchers. Reference pattern: one agent-loop contract, multiple frontends, capability policy, typed tool outcomes and independent verifier. Avoid claiming that Kiro's operational results are proven for 007.
Reactive agents provide a useful lifecycle question, not an adoption recommendation: suspend → external event → authorize → resume, with durable task/attempt/action identity, duplicate/out-of-order events, cancellation and fail-closed UNKNOWN. MCP Events is an experimental incubation; the repo currently labels Events Beta, not stable core MCP. Names/transport guarantees must be verified against a pinned revision before a wire-level PoC.
Ownership remains #42 action evidence / #43 persisted event reducer / #57 in-process loop. This issue adds prior art and desk-audit questions only. 007 is PARKED as product; no implementation, migration or autonomous runs authorized. If #42/#43/#57 already cover the required invariants, mark NO-CHANGE.