End-to-end validation surfaced that auto-dispatch parked every task at
"empty declared_scope": dispatch_node required plan["scope"], but the planner
emits only summary+phases (never a scope) and config.allowed_scope defaults to
None, so no node ever populated it. (The earlier dispatch test was
operator-initiated with an explicit scope; the auto planner->build->dispatch
path was never exercised until box-side App dispatch went live.)
Fix: when no planner-/operator-declared scope is present, dispatch_node derives
declared_scope from the candidate diff's own touched paths (ci_gate.diff_touched_paths).
This supplies the missing scope without relaxing any CI trust control — the
apply/verify workflow still INDEPENDENTLY re-checks the materialized diff against
the denylist + '..'-escape + this scope + the diff-hash binding, and the
agent-apply environment's required reviewer remains the human gate. A non-empty
diff that parses to zero touched paths still parks (fail closed).
Tests: the three old park-on-missing-scope cases now assert scope-from-diff
dispatch; added a park case for a diff with no parseable paths. 1527 pass.
The clarifier (ClaudeClarifier._turn) called claude_invoke with no max_turns,
inheriting the single-shot default (1). When the model's one turn did not
terminate in a final result the SDK raised 'Reached maximum number of turns (1)'
and, with no salvageable text, the call failed and crashed the clarify node —
leaving the task wedged at clarify with NO question posted to Slack (the human
never sees a clarifier prompt). Observed live on the R720.
Same single-shot flake the planner hit and fixed in PR #58 (_PLANNER_MAX_TURNS=4);
the clarifier never got the headroom. Give it the same: pass max_turns=4 (tools
stay off — still a fast reasoning->JSON completion).
- clarifier_llm.py: _turn passes max_turns=_CLARIFIER_MAX_TURNS (=4).
- tests: clarifier passes max_turns headroom to the invoke seam.
Full suite 1505 passed; ruff clean.
The real #60 root cause: failsafe_production_p3_wiring never bound a diff_builder,
so the build node fell back to builders.default_diff_builder (Claude). The Claude
agentic builder (max_turns + read-only tools) then exhausted its turn cap reading
files before it could emit a diff — 'Reached maximum number of turns (8)' live.
The intended builder already exists: builders_llm.as_diff_builder() routes the
build through DeepSeek fast_coder as a SINGLE model completion (with the
orchestrator's retrieval context + defensive _extract_diff + fail-safe no-op).
A single completion has no agent turn loop, so it CANNOT exhaust turns. Confirmed
get_fast_coder() works in the daemon environment.
- coordinator.py: failsafe_production_p3_wiring (P3-configured branch) now binds
builders_llm.as_diff_builder() as the diff_builder. Without it the build node
silently used the turn-exhausting Claude fallback.
- builders.py: default_diff_builder reverted to a safe SINGLE-SHOT, TOOL-LESS
fallback (max_turns=4, no allowed_tools/budget) — tools are what consumed the
turns; this fallback is no longer the live builder. Keeps _extract_unified_diff.
- tests: failsafe binds a real (callable) diff_builder when P3 is configured;
default_diff_builder is single-shot + tool-less.
Full suite 1504 passed; ruff clean.
The #60 max_turns/tools fix stopped the build-node crash but exposed the next
gap: with tools enabled the agentic builder reads files and its final text is
NARRATION (observed live: candidate_diff = 'Let me read the key source files to
get exact signatures bef…'), not a unified diff. That non-diff dispatched to CI,
could not be applied, and the task parked at verify.
- builders.py: default_diff_builder now extracts the unified diff from the
response via _extract_unified_diff — prefers a fenced ```diff block, else
slices from the first 'diff --git' header, dropping surrounding prose. A reply
with NO diff header raises BuildError so narration fails closed (the node fails
the task with a clear reason) instead of dispatching a bogus diff.
- builders.py: _render_build_prompt now instructs the model that its FINAL
message must be ONLY the unified diff in a single ```diff fenced block, no
narration before/after.
- tests: extract from fenced/bare diff with narration; reject narration-only
(BuildError); default_diff_builder returns the clean diff from a narrated
response and raises on a prose-only reply.
Full suite 1503 passed; ruff clean.
The Plane-2 builder (default_diff_builder) called claude_invoke with no
overrides, inheriting the subscription invoker's single-shot defaults
(max_turns=1, allowed_tools=[]). Diff synthesis is agentic, so the call
died with 'Reached maximum number of turns (1)' and every task failed at
phase=build.
- invoker.py: thread allowed_tools through subscription_invoker and
_collect_subscription_text (default None -> []), so callers can opt in;
single-shot reasoning nodes are unchanged.
- builders.py: default_diff_builder now passes max_turns=8, a read-only
tool allowlist (Read/Grep/Glob), and budget_usd=4.0. No write tools --
the builder returns the diff as data and performs no repo writes (D2/D11).
- Tests: builder agentic-config passthrough; invoker allowed_tools thread +
tool-less default guard (so future nodes must opt in explicitly).
Closes#60
_review_round_index counted EVERY review_verdicts entry, including the synthetic
human-gate verdict graph._apply_plan_decision folds in on a "request changes"
(reviewer == "human_plan_gate"). That inflated the count so a revised plan could
escalate prematurely without a fresh adversarial review.
Now counts only reviewer-authored verdicts: a new _is_reviewer_verdict excludes
entries tagged reviewer=="human_plan_gate" (read from the verdict dict's own
field — no graph.py import). A human request_changes now grants the revised plan
a fresh reviewer-round budget. Termination still bounded by MAX_PLAN_GATE_VISITS
(each request_changes consumes one gate visit). 1490 passed.
_revision_count read verdict.get("decision"), but verdicts are keyed "verdict"
(both reviewer and synthetic human-gate), so the count was always 0 and the
plan_node MAX_PLAN_REVISIONS self-park was dead code. Now reads "verdict" first
(fallback "decision"), matching _format_review_feedback's precedence; the
existing .strip().upper()==_REQUEST_CHANGES compare covers both request_changes
and REQUEST_CHANGES. Counts reviewer + human request_changes.
Cap composition: the review-loop round cap and MAX_PLAN_GATE_VISITS govern the
live loops; the planner MAX_PLAN_REVISIONS is now a correct backstop (was inert),
not a behavior change to the gate. Tests drive the real "verdict" key and prove
the previously-dead park fires. 1487 passed.
The planner's single-shot Claude call intermittently failed with "Reached
maximum number of turns (1)" — it needs slightly more headroom than the
clarifier to finish emitting its JSON. Phase A of the planner-reliability plan:
- plan_node now invokes with max_turns=4 (allowed_tools stays []; the extra
turns buy completion, not exploration).
- build_plan_prompt instructs the model to use no tools and return only JSON
(a tool_use would consume the single turn before the plan is emitted).
- plan_node auto-retries the model call exactly once on a TRANSIENT failure
(turn-cap exhaustion or an empty reply), and fails fast on DETERMINISTIC ones
(malformed JSON, missing/blank phases) — a retry would just reproduce those.
review_loop_llm.py is intentionally GPT-4.1 cross-family (no claude_invoke), so
it gets no turn-budget change. verifier_llm.py does use claude_invoke but is P3
build/verify scope — left for a follow-up.
Tests: max_turns passthrough; retry on turn-cap and on empty; no retry on
malformed JSON; the no-tools prompt line. 1387 passed.
High-recall /sh-security-review fan-out + proof-or-kill verifier found two
confirmed HIGH; both now closed (verified empirically against the working tree):
- LOGIC-RACE-01 (HIGH, CWE-835): the build-loop budget was structurally dead
(verifier read a shared wiring-time VerifierConfig.build_loops, always 0, so
the max_build_loops park never fired -> a perpetually-failing task looped
BUILD->DISPATCH->VERIFY forever, force-pushing + firing a CI run each round).
Threaded build_loops through durable PipelineState/TaskRecord; verifier reads
state.get('build_loops',0), writes the incremented count back on each FAIL, and
PARKS at max_build_loops. Parks after exactly N failures, never unbounded.
- SEC-01 (HIGH, CWE-532) + SEC-02 (MED, CWE-214): p3_rollback.sh echoed the live
App JWT to stdout in default dry-run and passed it as a gh argv literal. Added
redact_secrets (Bearer/Authorization/ghX_/PEM masking) through run_or_plan; the
App uninstall now uses curl -H @<0600 tempfile> (JWT never on argv), shredded
after. Empirical: app/incident/all dry-runs leak 0 JWT occurrences.
- SEC-03 (MED, CWE-798): assert_no_write_token now applies the PEM regex + the
configured App-ID to env/config VALUES (not just files) — an App private key
under a benign env name is caught.
- SEC-04 (LOW) + P3-IAC-08 (LOW): tightened the box GITHUB_TOKEN fallback /
value-scan; staged-only WARN on the live workflow revert.
Suite: 1382 passed, ruff clean. Branch only; not merged/deployed.
NOTE: re-verifier flagged SEC-01 as open by grepping COMMITTED blobs (the fix was
uncommitted working-tree state); independently confirmed closed empirically.
Wire the box-side build->dispatch->verify run identity so the verifier gate
can bind to the CI run the dispatcher triggered:
- task_model: add run_id / ci_correlation_tag / dispatched_at to TaskRecord +
PipelineState (+ dict round-trip).
- dispatcher: RunLocator seam + DispatchResult; dispatch_apply_verify stamps a
dispatched-at watermark, fires, then resolves the run via the workflow
run-name (gh run list; the per-task_id concurrency group makes it
unambiguous). Fails closed to run_id=None.
- dispatch_invoker: persist run_id/dispatched_at/ci_correlation_tag into state.
- workflow: additive run-name surfacing inputs.task_id as the correlation key
(flagged for the C1 /sh-security-review + GPT-4.1 cross-review re-run).
- docs: P3-PHASE0-DESIGN.md records the async-resume design decision.
Part of Phase 0 (feat/agent-team-p3-box-integration). No behavior change on the
default path: P3 wiring is still opt-in/inert.
Two root causes behind 'every task parks, and slowly':
1. SINGLE-SHOT INVOKER: subscription Claude calls ran as 40-turn, tool-enabled
agentic sessions (--max-turns 40, $2 budget) for what are pure reasoning->JSON
completions — minutes-long, and Claude wandered/returned unparseable output.
_DEFAULT_MAX_TURNS 40->1 + allowed_tools=[] -> fast deterministic single turn.
2. PLANNER FEEDBACK KEY MISMATCH: the review stage writes verdict/findings, but
_format_review_feedback read decision/notes/comment (never present) -> the
planner re-planned with EMPTY feedback, re-introduced the rejected flaw
('assumptions persist'), hit the review cap, parked. Now reads verdict/findings
(old keys kept as fallback) so GPT-4.1's objections reach the re-plan.
Plus: park notifications infer the phase the task was IN (review/plan/clarify)
instead of the terminal 'parked'.
Tests: planner real-verdict-keys regression + coordinator phase-inference. 1149 pass.
Gap-audit findings:
- G1 (blocks-feature): make_cross_reviewer_invoker / make_fast_coder_invoker did
'from models import' without putting the orchestrator root on sys.path. The
run-team serve daemon only bootstraps agent-team/, so on the live box every
GPT-4.1 plan review hit ModuleNotFoundError -> review_plan's blanket except
silently fail-closed to REQUEST_CHANGES (GPT-4.1 never actually ran). Both
in-process invokers now call invoker_multi._ensure_orchestrator_on_path()
before the deferred import. WS1 introduced this when it swapped the review
default from the subprocess invoker to in-process.
- G4 (degrades): the handbook context_provider was wired into the planner only;
default_clarify_node_factory now accepts + forwards it, and run-team wires it
into build_clarify_node too, so clarifying questions are handbook-aware.
Tests: +2 regression tests (path-bootstrap, clarifier threading); _FakeCoordinator
gains build_clarify_node. 1142 passed, ruff clean.
CI has no fastapi (box-only dependency); api.py imports it lazily. Guard the
7 TestClient smoke tests with skipif(find_spec('fastapi') is None) so they
skip in CI instead of failing collection, leaving the 14 invoker_multi tests
running. Also apply ruff format to the 5 WS1 files CI flagged.
- invoker_multi.py: multi_invoke(prompt, *, model) dispatches to GPT-4.1
(cross_reviewer), DeepSeek (fast_coder), or Gemini (scanner) in-process;
bind_multi_invoker() wires the review seam; lazy models import
- review_loop_llm: swap default_plan_reviewer from make_run_py_invoker()
to make_cross_reviewer_invoker() (in-process GPT-4.1); subprocess path
retained as make_run_py_invoker() for opt-in use
- builders_llm: add make_fast_coder_invoker() in-process DeepSeek path;
rename subprocess path to subprocess_build (opt-in fallback); default_build
now delegates to make_fast_coder_invoker()
- api.py: FastAPI app with bearer-token auth (AGENT_TEAM_API_TOKEN env),
POST /tasks, GET /tasks/{id}, POST /orchestrator/invoke; binds 127.0.0.1;
code only — not started here
- run-team.py: bind_multi_invoker() called in serve() alongside
bind_subscription_invoker()
- tests: 21 new WS1 tests + updated builders_llm tests (1065 total, all passing)
- retriever: add save_memory() writing to _box-drafts/ review queue, add
memory_dir param to retrieve() for isolated test routing
- handbook: new load_handbook_conventions() with safe no-op contract (returns
"" when dir absent/empty/unreadable, never raises)
- clarifier_llm: add context_provider=None seam to ClaudeClarifier and
build_claude_clarifier_callables(); failure in provider is silent
- planner: add context_provider=None seam to build_plan_prompt() and plan_node()
- coordinator: thread context_provider through default_plan_node_factory(),
only pass kwarg when non-None to preserve stub-monkeypatching in tests
- tests: 19 new WS5 tests covering all seams (1063 total, all passing)
* feat(agent-team): Plane-1 Tier-3 fixer — dependency-cve finding -> patch + CI dispatch (opt-in/inert)
The fixer (design §4 fixer row, §7 Phase 5, §3.3.2) takes a CONFIRMED,
low-risk dependency-cve finding (the narrowest fix class) and produces:
* a fix SPEC (Claude, via the §3.1 billing seam), and
* a minimal bump PATCH (DeepSeek fast_coder, via the orchestrator run.py
path that builders_llm uses),
records the candidate diff + its content-hash, and emits the org-CI
workflow_dispatch inputs (task_id / diff_artifact_name / expected_diff_hash /
declared_scope) for the gate-passed P3-live apply/verify surface.
INERT / opt-in / fail-safe, mirroring build_verify_wiring:
* plan_fix dispatches NOTHING; dispatch_fix has NO default dispatcher
(the box holds no write token, D2) so an un-wired call can never fire a
workflow.
* no git/patch/subprocess/fs-write in executable code — the patch is emitted
as diff TEXT only; CI applies it and opens a DRAFT PR, the box never
applies/pushes/merges.
* untrusted-patch hygiene: the generated diff is confined box-side to the
single dependency manifest (declared_scope) and rejected via
ci_gate.denylist_violations if it escapes scope or touches the
trust-control surface — defense-in-depth with the CI guard.
* bad/ambiguous findings (wrong check/status/category, missing
package/fixed_version, ambiguous fixed_version, unparseable/empty diff)
yield a FAILED no-op plan, never a fabricated fix.
29 new pytest tests under agent-team/tests/test_fixer.py.
* feat(agent-team): run-team.py 'fix --dry-run' subcommand for the Plane-1 fixer
Adds the fixer front door to the operator CLI: load one confirmed
dependency-cve finding from a dependency-cve.json report (--report
--finding-id), plan the fix, and in --dry-run print the spec + patch + the
org-CI workflow_dispatch inputs WITHOUT dispatching anything.
Opt-in/inert: the command binds NO workflow dispatcher and holds no write
token, so even an ok plan only prints; live dispatch is provisioning-gated
(refuses to run without --dry-run). A non-fixable finding prints the
fail-safe reason and exits 1.
4 new pytest tests under agent-team/tests/test_run_team.py.
clarifier_llm: sha1 -> sha256 for the non-security cache-discriminator (CWE-327 false positive). github_adapter + github_intake: inline nosemgrep on the urlopen lines (dynamic-urllib-use-detected) — the URL is built from a fixed https GitHub API base, dynamic part is the path only, no SSRF/file:// surface (extends the existing noqa:S310 trusted-host judgment to semgrep). Scanner now reports 0 mediums on the agent-team scope.
build_verify_subgraph: BUILD->VERIFY nodes + route_after_verify. build_graph gains an opt-in build_verify param that repoints the review 'build' route at the subgraph (BUILD->VERIFY->{approved->END | loop->PLAN | parked->END}); default unchanged (P2). Coordinator build_verify_wiring composes it INERT (no ci_result -> ci_gate BLOCK -> PARKED; LLM is fix-proposer only, never declares green). NOT enabled in production: the live CI apply/verify + OIDC stays held for its /sh-security-review + GPT-4.1 cross-review gate. Also escapes untrusted intake text in logs (log-injection hygiene).
review_loop_llm -> GPT-4.1 cross_reviewer (orchestrator run.py); builders_llm -> DeepSeek fast_coder (INERT, proposes diff text only); verifier_llm -> ci_gate is sole PASS authority, Claude is fix-proposer only. Hardens review_loop.parse_verdict to word-boundary matching, adds a fail-closed subprocess timeout, and bind_review_node (single-arg, no LangGraph config injection). All fail safe on untrusted model output.
Adds the billing-seam invoker (claude_agent_sdk subscription-OAuth, deferred import, API/Bedrock paths) and the Claude-backed clarifier callables (ConfidenceAssessor/QuestionGenerator, one call/turn memoized on (thread_id,len,content-hash), fail-safe to 0.0 so garbage never clears the 98% human gate).
Addresses the confirmed findings from /sh-security-review + the GPT-4.1
cross-review of the Plane-2 scaffold. Full suite: 589 passed; ruff clean.
FIXED (proven-exploitable):
- CI-guard denylist bypass (HIGH): Python fnmatch '**/' is non-recursive, so
root-level template.yaml/*.tf/cdk.json/*.pem/*.key/*-stack.* evaded the
trust-control surface. Replaced fnmatch with a recursive, case-insensitive
glob->regex matcher. (verified: fnmatch('template.yaml','**/template.yaml')==False)
- CI-guard scope bypass (HIGH): a '**' declared_scope made every path in-scope.
Scope is now concrete-prefix confinement (reduces a glob to its leading
metacharacter-free segments; '**' -> empty -> dropped -> unscoped reject).
- Box-side vs CI denylist divergence (MED): builders.py _DENY_PATTERNS now covers
Terraform, *.pem/*.key, CDK stack files, .github/actions, *iam*, bare policy*.json
(case-insensitive), matching the CI surface.
- force-resume was backwards (MED): it superseded the answered row recovery
resumes from, making a stuck task permanently un-resumable while printing
success. Now re-opens an EXPIRED (parked) question via a new reopen_question
CAS helper; never supersedes an answered row; honest exit codes.
- operator attribution (MED): run-team.py --operator defaulted to "" -> now the
OS login, so destructive actions are always attributable.
- audit-log append race (MED): replaced read-modify-rewrite (lost records under
concurrent operators) with an O_APPEND single-line write, mode 600 enforced.
- lstrip("ab/") path-mangling in the symlink error path -> regex prefix strip.
Regression tests added across test_ci_gate_workflow / test_builders / test_run_team
/ test_schema. Design-level findings (resume-worker durability, egress breadth,
answered_at ordering, DB-swap TOCTOU, diff-hash threat-model) are pre-deployment
/ P1-build-proper and recorded with written justification in
agent-team/.security-review/suppressions.json; CI README diff-hash wording made
honest.
Resolves three execution-proven verifier findings from the scaffold review.
Full suite: 548 passed, 1 skipped (stable across repeated runs); ruff clean.
builders denylist (§3.3.2 #2): scan was +++-only and missed header-only
sections. Now section-driven off `diff --git a/<src> b/<dest>`, catching the 4
proven bypasses — delete of a denied path, mode-change-only, `copy to` a denied
path, out-of-scope delete (regression tests for each).
§3.3.1 compare-and-set concurrency: BEGIN IMMEDIATE moved inside guarded retry;
each CAS now runs on its own connection (shared sqlite3.Connection cannot hold
two transactions, and is unsafe for concurrent use even for reads). connect()
stashes the db path on a Connection subclass so the path is derived by a
thread-safe attribute read, not a PRAGMA on the shared conn; busy_timeout set
before the WAL pragma. Added shared-connection concurrent regression tests
(distinct + same question) — previously raised "transaction within a
transaction".
operator CLI (run-team.py): added the design-named re-deliver and force-resume
verbs (were missing); audit now records the attempt BEFORE the mutation and the
outcome after, so a ledger mutation can never land without a trail; main()
catches OSError instead of leaving an uncaught traceback on audit-write failure.
