Repository navigation
Restore green main: CI history, environment-independent tests, honest receipts, 2026-10-09 roadmap review and owner decisions - #76
Merged
Conversation
…provement plan Add docs/reviews/roadmap-review-and-improvement-plan-2026-10-09.md, a dated review of the repository state with a proposed, gated Phase 0-4 plan: - Red main: CI runs 851-858 fail on `test (ubuntu-latest, 3.12)` and `acceptance-all`. The 9 failing tests were re-run off-runner: 8 of 9 reproduce in a shallow clone, 2 of 9 with full history (`test_real_delivery_ledger_passes` reads gitignored `.teaagent/delivery/`; `test_rollback_refuses_when_head_not_on_sandbox_branch` assumes `git init` creates `main`). One failure (`'[exit '` receipt fragment) is unreproduced. This supersedes the "not yet reproduced off-runner" wording in the CI debt register; the register itself is left for the Phase 0 fix commit. - Overdue dated decisions: ADR-0031 (2026-09-29) has no recorded disposition; DR-006 falsifier review and T5 survey have no record; ADR-0043 disposition stays 2026-12-09. - Evidence starvation: G1 unexercised, B1 pending, no owner friction entry since 2026-06-22. - Owner asks with defaults, critical-thinking check, falsifiers. Also: index the record in docs/INDEX.md, add a dated review line to docs/roadmap-status.md (no horizon, milestone, goal, or gate status moved), restate the G5 corpus figures from the regenerated inventory (668 files / 667 inventoried / 8 constitution / 369 working / 290 archive), register G-P2-21 in the action register, and regenerate the docs inventory, aging dashboard, and OKF bundles. The release docs evidence bundle is rebuilt on the clean tree in the follow-up commit per the bundle protocol. Action: G-P2-21 Roadmap-Status: unchanged Constraint: documentation only; no runtime, test, CI, or gate change; all plan rows are Proposed Tested: bash scripts/verify_docs.sh (all checks pass except the evidence bundle, which the bundle protocol rebuilds on a clean tree next); pytest tests/test_docs_consistency.py tests/acceptance/test_docs_acceptance_count_accuracy.py (56 passed); scripts/check_action_register_link.py exit 0; scripts/run_test_tier.py --tier smoke (203 passed, 1 skipped) Not-tested: full unit suite with coverage; CI run on this branch Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…0-09 roadmap review)
Regenerate docs/generated/release-docs-evidence.{md,json} on the clean tree
at 1c8b5af and refresh the docs inventory hash for the regenerated bundle.
The recorded gate run is green (validate_docs_consistency.py in full
test-quality mode, report_docs_aging.py --check).
Action: G-P2-21
Roadmap-Status: unchanged
Constraint: generated artifacts only; no hand edits
Tested: python3 scripts/build_release_docs_evidence_bundle.py on a clean tree with dev extras installed (gate pass); bash scripts/verify_docs.sh after commit
Confidence: high
Agent: Claude Code
Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…ts depending on local state Phase 0 (R0-1, R0-3, R0-4, R0-7) of the 2026-10-09 roadmap review: - .github/workflows/ci.yml: `fetch-depth: 0` on the `test` matrix and `acceptance-all` checkouts. Six of the nine red-main failures (`test_real_history_87d1c61_passes`, four `test_refresh_competitive_docs` tests, `test_check_docs_aging_dashboard_passes_for_repo`) need git history that a shallow checkout does not have; the 2026-09-29 fixes added it only to `use-case-matrix`, `lint`, and `review-institution`. - tests/test_g2_g3_g22_focused.py: pin the initial branch with `git symbolic-ref HEAD refs/heads/main` after `git init`; the test ran `git checkout main` and GitHub runners default to `master`. - tests/test_evidence_ledger.py: `test_real_delivery_ledger_passes` skips with a reason when no `.teaagent/delivery/` ledger exists (the directory is gitignored, so the test could only pass on the author's machine); validation still runs when ledgers are present. No assertion weakened. - scripts/run_acceptance_tier.py: run pytest with `sys.executable` instead of a literal `python3` so the tier runner uses the interpreter that launched it. Action: G-P2-21 Roadmap-Status: unchanged Constraint: CI configuration and test-environment fixes only; no teaagent/ runtime change; no assertion weakened or removed Tested: pytest tests/test_g2_g3_g22_focused.py (4 passed with init.defaultBranch unset, =master, =main); pytest tests/test_evidence_ledger.py -rs (8 passed, 1 skipped with reason); pytest tests/test_acceptance_tier_runner.py (3 passed); ruff check + format --check on changed Python files; YAML parse of ci.yml Not-tested: CI run on this branch (follows on push) Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
Continuous-ratchet item from the 2026-10-09 roadmap review: count `Any` annotations under teaagent/ (ignoring comment text and typing import lines), fail when the total exceeds ANY_BASELINE = 2177 (measured at 0d9b4ae), and print a HINT to lower the baseline when the count drops. The baseline can only be lowered; `--update-baseline` prints the edit and never writes. The CI `lint` step "Any-count baseline (typing-debt ratchet)" was already committed in 0b227fc (it was staged together with the fetch-depth change); this commit adds the script and its contract tests so that step can run. Action: G-P2-21 Roadmap-Status: unchanged Constraint: stdlib-only script under scripts/ plus tests; no teaagent/ runtime change; CI step is additive and reversible Tested: pytest tests/test_check_any_count.py (4 passed); ruff check + format --check; mypy on both files (no issues); python3 scripts/check_any_count.py --fail-on-increase (OK: 2177 <= 2177) Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
… failing Phase 0 (R0-2) of the 2026-10-09 roadmap review. Six tests ran the real repository docs-aging check or read a specific commit and therefore failed on any shallow checkout (`shallow clone: docs aging check requires full git history`, `unknown revision 87d1c61`). They now skip with an explicit reason when the checkout cannot support them and still run in full on a complete clone; no assertion was changed. - test_support.py: `repo_is_shallow()` (git rev-parse --is-shallow-repository) and `commit_exists(sha)` (git cat-file -e <sha>^{commit}); git errors count as "not shallow" / "absent". - tests/test_refresh_competitive_docs.py (4 tests), tests/test_report_docs_aging.py (1 test): skip when the repo is shallow. - tests/test_check_dr006_gate_trailer.py: `test_real_history_87d1c61_passes` and `test_real_history_12e7148_fails` skip when the commit is absent (the latter previously passed vacuously on a shallow clone because the script fails on an unknown revision for the wrong reason). Action: G-P2-21 Roadmap-Status: unchanged Constraint: test and test-support changes only; skips carry reason=; no production code or assertion changed Tested: full checkout pytest on the three files (25 passed, 0 skipped); `git clone --depth 1` simulation with the edited files copied in (19 passed, 6 skipped, 0 failed; the untouched HEAD versions gave 6 failed); ruff check + format --check; mypy on the four files (no issues) Not-tested: the 12e7148 guard on a shallow clone (same mechanism as the 87d1c61 guard) Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…y step The ninth red-main failure (`test_real_run_receipt_completeness_from_plan`, missing receipt fragment `'[exit '`) is reproduced off-runner by removing `rg` from PATH: the fake-adapter verify step runs `rg`, GitHub ubuntu-latest ships no ripgrep, `workspace_run_shell_inspect` raises FileNotFoundError, the runner records `tool_call_failed`, the run still completes, and the receipt renders the command with no exit marker. The earlier first-run-banner lead was wrong. This commit only removes the environment dependency: the verify command is `grep 'return a + b' calc.py` (BRE, `+` literal; grep is in the inspect allowlist). No assertion changed. The receipt-honesty gap (a failed command rendering silently) is fixed separately in the receipt builder. Action: G-P2-21 Roadmap-Status: unchanged Constraint: test input change only; assertions unchanged; no teaagent/ change Tested: pytest tests/acceptance/test_run_evidence_summary_flow.py under env -i with a PATH lacking rg plus HOME=$(mktemp -d) CI=true GITHUB_ACTIONS=true (11 passed); ruff check + format --check Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…ilence Governance gap found while root-causing the ninth red-main failure: when a shell inspect call fails to start (e.g. the executable is missing), the runner records `tool_call_started` then `tool_call_failed` and the run still ends `completed`, but `extract_commands_run` ignored `tool_call_failed` and `format_run_receipt` only rendered a suffix when an exit code existed. The receipt therefore showed ` - <command>` with no outcome, so a verification step that never ran read like one that had. - teaagent/run_evidence.py: `CommandEvidence.error` (optional); the bundle serializes it; `extract_commands_run` consumes `tool_call_failed` matched by call_id, keeps `exit_code=None`, records the classified error text. - teaagent/run_receipt.py: `[exit N]` unchanged; `[failed: <error>]` when a command errored without an exit code; `[outcome unknown]` when it only started. Never emits a fake `[exit ` for a command that did not run, so `check_receipt_completeness` still flags such receipts. - tests: three evidence tests and three receipt tests pin the three branches. Residuals recorded, not fixed here: `auto_derive_known_gaps` still keys on non-zero exit codes only; the runner's OUTCOME_UNKNOWN `tool_call_failed` without a prior start has no command text and stays out of the receipt list. Gate: governance-gap Action: G-P2-21 Roadmap-Status: unchanged Constraint: receipt and evidence fold only; no runner, audit emission, or schema change; no existing assertion changed; success-path output identical Tested: pytest tests/test_run_evidence.py tests/test_run_receipt.py tests/acceptance/test_run_evidence_summary_flow.py tests/test_evidence_completeness.py tests/acceptance/test_five_minute_proof_flow.py tests/acceptance/test_run_undo_acceptance_flow.py (75 passed); ruff check + format --check; mypy on both modules (no issues); validate_runner_invariants.py, validate_event_spine_wiring.py, check_audit_schema_conformance.py, check_god_modules.py all exit 0 Not-tested: full suite (running separately); live provider runs Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…ified root causes - docs/governance/testing-standards.md: replace the 2026-09-30 "not yet reproduced off-runner" wording with the verified per-test cause table (six shallow-history tests, one gitignored-state test, one default-branch assumption, one missing `rg` plus a receipt-honesty gap) and the repairs that landed (R0-1..R0-5, R0-7). Adds the rule derived from the incident: tests must create repo/history/HOME state under tmp_path or skip with a reason. - docs/reviews/roadmap-review-and-improvement-plan-2026-10-09.md: row 9 and R0-5 restated from "unreproduced, first-run banner lead" to the reproduced cause (ripgrep absent on GitHub ubuntu-latest; `tool_call_failed` dropped by the receipt builder); the critical-thinking section records the falsified inference. - Regenerated docs inventory and knowledge bundles. The release docs evidence bundle is rebuilt on the clean tree in the follow-up commit. Action: G-P2-21 Roadmap-Status: unchanged Constraint: documentation only; no status, gate, or horizon moved Tested: bash scripts/verify_docs.sh (all checks pass except the evidence bundle, rebuilt next per the bundle protocol); pytest tests/test_docs_consistency.py tests/acceptance/test_docs_acceptance_count_accuracy.py (56 passed) Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…0 repairs)
Regenerate docs/generated/release-docs-evidence.{md,json} on the clean tree
and refresh the docs inventory hash for the regenerated bundle. The recorded
gate run is green (validate_docs_consistency.py in full test-quality mode,
report_docs_aging.py --check).
Action: G-P2-21
Roadmap-Status: unchanged
Constraint: generated artifacts only; no hand edits
Tested: python3 scripts/build_release_docs_evidence_bundle.py on a clean tree with dev extras installed (gate pass); bash scripts/verify_docs.sh after commit
Confidence: high
Agent: Claude Code
Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…ification Addendum §7.1 to the 2026-10-09 roadmap review: which Phase 0 rows landed in which commits, and the full unit suite result on this tree (27 failed / 6,736 passed / 27 skipped) classified per failure against an origin/main worktree: 15 provider-DNS, 3 ssh-keygen, 4 root-bypasses-chmod, 2 container SSL env, 3 pre-existing flake candidates (cooldown wall-clock sleep, Hypothesis deadline, one xdist-only failure). None touches the Phase 0 files. Action: G-P2-21 Roadmap-Status: unchanged Constraint: documentation only Tested: bash scripts/verify_docs.sh (all checks pass except the evidence bundle, rebuilt next per the bundle protocol) Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…0 execution record)
Regenerate docs/generated/release-docs-evidence.{md,json} on the clean tree
and refresh the docs inventory hash for the regenerated bundle. The recorded
gate run is green (validate_docs_consistency.py in full test-quality mode,
report_docs_aging.py --check).
Action: G-P2-21
Roadmap-Status: unchanged
Constraint: generated artifacts only; no hand edits
Tested: python3 scripts/build_release_docs_evidence_bundle.py on a clean tree with dev extras installed (gate pass); bash scripts/verify_docs.sh after commit
Confidence: high
Agent: Claude Code
Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…ands for real `TestShellCommandInjectionFix::test_shell_blocks_dangerous_commands` called `run_shell()` with `rm -rf /`, `mkfs`, and `dd if=/dev/zero of=/dev/sda` and relied on the host to refuse them. `run_shell` has no command denylist by design (governance lives in ApprovalPolicy), so the commands were executed: run as root in a container during the 2026-10-09 full-suite run, `dd` wrote zeros to a device-named path for the full 30 s timeout. On a developer machine with write access to the block device this test would destroy the disk. The test only "passed" on CI because the runner cannot write /dev/sda. Replacement `test_shell_dangerous_commands_are_governed_not_executed`: - monkeypatches `teaagent.workspace_tools._shell.subprocess.run` to raise, so the test fails loudly if any process is ever spawned; - asserts both shell tools are registered `destructive=True, read_only=False`; - asserts `ApprovalPolicy` raises `ToolPermissionError` for each dangerous command in read-only, workspace-write, and unapproved prompt mode. No other test in the suite executes a destructive command through a real subprocess (swept `tests/` for `rm -rf /`, `mkfs`, `of=/dev/sd`; remaining hits are classifier/policy inputs only). Action: G-P2-21 Roadmap-Status: unchanged Constraint: test-only change; asserts the real governance boundary instead of host behavior; no production code change Tested: pytest tests/test_security_fixes.py (all passed, no subprocess spawned); ruff check + format --check Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…ity test `build_workspace_tool_registry` takes a root path, not a WorkspaceToolConfig; the previous commit passed the config and the new test raised TypeError. The earlier commit's "all passed" claim was wrong (the summary line was read through a pipe that masked the exit code); this commit records the corrected result. Action: G-P2-21 Roadmap-Status: unchanged Constraint: test-only one-line fix Tested: pytest tests/test_security_fixes.py exit 0 (20 passed, exit code checked directly); ruff check + format Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…n fixture `check_provider_connectivity` resolves `api.openai.com` for remote providers; in any sandbox without that DNS, `preflight`/`daily`/`plan` report `ready=False` and exit 2, and 15 tests that assert readiness output (not DNS) fail with `assert 2 == 0`. CI passes only while the runner resolves the host. - tests/conftest.py: opt-in (not autouse) `offline_provider_connectivity` fixture monkeypatches `teaagent.preflight.check_provider_connectivity` to report OK, and also patches any teaagent module that imported the symbol by name (none today; the scan reads `__dict__` so the package's lazy `__getattr__` never runs). - Applied to 12 function-style tests in test_preflight.py, acceptance/test_context_pack_read_only_flow.py, test_daily_cli.py, test_daily_tui.py, test_first_hour_e2e_flow.py, test_repo_map_quality_large_repo_flow.py. The three `unittest.TestCase` methods (test_tui.py, test_real_usage_agents.py) follow in a separate commit because those files are being edited for root pollution as well. - No test asserts the DNS-failure message; `tests/test_dogfood_first_run_g23_g38.py` (fake provider skips DNS) still passes. No production code changed. Action: G-P2-21 Roadmap-Status: unchanged Constraint: test-only; fixture is opt-in; no assertion changed; teaagent/preflight.py untouched Tested: in this container (no DNS for api.openai.com) the six files plus test_dogfood_first_run_g23_g38.py pass (exit 0, checked directly); before the fixture the same 15 tests failed; ruff check + format --check Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…clock and deadline flakes
Follow-ups from the 2026-10-09 test-suite state-dependence sweep:
- chmod-based PermissionError tests skip when running as root (root ignores
file modes): test_ws3_compliance_audit.py (2), integration/test_config_loader.py
(1), test_preflight_env_health.py (1).
- ssh-keygen presence: acceptance/test_security_ssh_signatures_flow.py skips
the three tests that generate a keypair when `shutil.which('ssh-keygen')`
is None (previously FileNotFoundError outside the try/skip);
test_strategic_features.py replaces the POSIX-only `which` subprocess with
`shutil.which`.
- Flakes seen under xdist load: test_hypothesis_invariants.py
`TestApprovalHashExactness` property tests get
`@settings(deadline=None, suppress_health_check=[HealthCheck.too_slow])`;
test_audit_health.py `test_assess_cooldown_expired` drives
`teaagent.audit.time.monotonic` through a fake clock instead of sleeping
2 ms past a 1 ms cooldown (audit.py uses only `time.monotonic`).
- Hygiene: test_governance_adversarial_runtime.py writes its guard-regression
marker under the system temp dir with a unique name instead of a fixed
/tmp path; test_intent.py passes `--root tmp_path` so the clarify run no
longer writes `.teaagent/welcomed` into the repository checkout.
No assertion changed; every skip carries reason=. The unused
`_make_conflicting_branches` helper deletion in test_real_usage_tui.py
lands with that file's root-pollution fix in the next commit.
Action: G-P2-21
Roadmap-Status: unchanged
Constraint: test-only; no production code change
Tested: pytest --random-order on the nine files (exit 0, checked directly); as root here 8 skipped with reasons; test_hypothesis_invariants.py and test_audit_health.py three --random-order runs each green; ruff check + format --check
Confidence: high
Agent: Claude Code
Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…S patch for TestCase methods Tests that built `TeaAgentTUI(...)` or ran the CLI without a root wrote runtime state (`runs/*.jsonl`, `suspension-*.json`, `sessions/*.json`, `seen_version`) into the repository checkout's own `.teaagent/`, the same directory used as owner-use evidence (the 2026-09-17 review fixed one such leak; this closes the rest found by the sweep). - tests/test_tui.py: `TUITests.setUp` creates a TemporaryDirectory and all 72 `TeaAgentTUI(` constructions pass `root=self.tui_root`. - tests/test_real_usage_tui.py: `_ScratchRootMixin` for the nine scenario classes; 26 constructions get `root=`; the unused `_make_conflicting_branches` helper (assumed `git init` creates `main`) is deleted. - tests/test_real_usage_agents.py: same mixin; 8 CLI invocations pass `--root`; `model smoke` is left alone because its parser rejects `--root`. - The three `unittest.TestCase` readiness tests (`test_tui_preflight_command_uses_current_settings`, `test_b1_cli_preflight`, `test_b3_cli_plan_mode`) patch `teaagent.preflight.check_provider_connectivity` via `unittest.mock.patch` with `addCleanup`, completing the DNS-independence change from 4399a1c. Measured leak on clean HEAD: 1 test / 3 files (test_tui.py), 3 tests / 3 files (test_real_usage_tui.py), 1 file (test_real_usage_agents.py). After: zero new files under the repo `.teaagent/` when running the three files. No assertion added or removed. Action: G-P2-21 Roadmap-Status: unchanged Constraint: test-only; no production code change; live-test gate untouched Tested: pytest tests/test_tui.py tests/test_real_usage_tui.py tests/test_real_usage_agents.py (274 passed, 6 skipped, exit 0 checked directly) with a marker-file probe showing 0 new files under .teaagent/; ruff check + format --check Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…tructive-test incident Addendum §7.2 to the 2026-10-09 roadmap review: the safety incident in tests/test_security_fixes.py (destructive shell commands executed for real), the DNS/root/tool-presence/flake/scratch-root repair commits, the final full-suite result on this tree (2 failed / 6,760 passed / 34 skipped, both failures container SSL env), and the residual items recorded but not fixed. Action: G-P2-21 Roadmap-Status: unchanged Constraint: documentation only Tested: bash scripts/verify_docs.sh (all checks pass except the evidence bundle, rebuilt next per the bundle protocol) Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…follow-up record)
Regenerate docs/generated/release-docs-evidence.{md,json} on the clean tree
and refresh the docs inventory hash for the regenerated bundle. The recorded
gate run is green (validate_docs_consistency.py in full test-quality mode,
report_docs_aging.py --check).
Action: G-P2-21
Roadmap-Status: unchanged
Constraint: generated artifacts only; no hand edits
Tested: python3 scripts/build_release_docs_evidence_bundle.py on a clean tree with dev extras installed (gate pass); bash scripts/verify_docs.sh after commit
Confidence: high
Agent: Claude Code
Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
… keep startless failures visible Closes the two residuals recorded in c6b57f6: - `auto_derive_known_gaps` now emits `command_failure` ("Command did not complete: <cmd> (<error>)") for a command that errored before an exit code, and `command_outcome_unknown` ("Command outcome unknown: <cmd>") for a command that only started. Non-zero-exit handling, ordering, and dedup are unchanged. - `extract_commands_run`: a `tool_call_failed` for a recognized shell tool that carries no command text (the runner's OUTCOME_UNKNOWN path has no arguments and no prior start) gets the placeholder command `<tool_name: no command recorded>` so the receipt renders `[failed: ...]` instead of dropping the entry. Non-shell tools keep today's filter. - Five tests pin the gap strings, the placeholder, the non-shell filter, and the unchanged success path. Gate: governance-gap Action: G-P2-21 Roadmap-Status: unchanged Constraint: evidence fold and known-gap derivation only; no runner, audit emission, or schema change; no existing assertion changed Tested: pytest tests/test_run_evidence.py tests/test_run_receipt.py tests/acceptance/test_run_evidence_summary_flow.py tests/test_evidence_completeness.py tests/acceptance/test_five_minute_proof_flow.py tests/acceptance/test_run_undo_acceptance_flow.py tests/test_run_evidence_summary.py (97 passed); ruff check + format --check; mypy teaagent/run_evidence.py (no issues); validate_runner_invariants.py, validate_event_spine_wiring.py, check_god_modules.py exit 0 Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…; archive-growth signal Sweep residuals (all test-only unless noted; no assertion weakened): - tests/test_llm_internals.py: `TestBuildSslContextFromEnv` clears the four env vars `build_ssl_context_from_env` reads (REQUESTS_CA_BUNDLE, SSL_CERT_FILE, TEAAGENT_TLS_CLIENT_CERT/KEY) so hosts that export a CA bundle no longer fail the two tests. - tests/test_real_usage_scenarios.py, tests/test_real_usage_agents.py: the live-gated key helper reads the repo-root env file, never `Path.cwd()`. - tests/test_refresh_competitive_docs.py: snapshot/restore of the two tracked generated docs the refresh script rewrites, so the file leaves `git status` clean. Finding recorded, not fixed: `check_competitive_docs --check` also writes those files despite its docstring. - Loopback guard (`can_bind_loopback`) on six acceptance files that start a local HTTP server without one. - tests/test_preflight.py: explicit DNS-failure-path test for `check_provider_connectivity` (gaierror → False with "could not be resolved"; `fake` never touches DNS). - .github/workflows/ci.yml: `timeout-minutes` 30 on `test`, 20 on `acceptance-all` (soft budget 20 min per the 2026-10-09 review). - scripts/report_docs_aging.py: "Archive-tier growth (files added per month, last 6 months)" table in the G5 corpus-cost section, from one `git log --diff-filter=A --no-renames` pass; two tests; dashboard and inventory regenerated. Action: G-P2-21 Roadmap-Status: unchanged Constraint: tests, CI config, and docs tooling only; no teaagent/ runtime change Tested: pytest on all 14 touched test files plus the evidence flows (256 passed, 7 skipped after regenerating the dashboard; the single prior failure was the stale dashboard); ruff check + format --check (16 files); mypy scripts/report_docs_aging.py (no issues); report_docs_aging.py --check passes; YAML parse of ci.yml Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…s addCleanup `_ScratchRootMixin.setUp` calls `self.addCleanup`, which mypy could not see on a bare mixin (2 errors on the full `mypy teaagent/ tests/` run). Deriving the mixin from `unittest.TestCase` resolves it; the underscore name keeps pytest from collecting the mixin itself. Action: G-P2-21 Roadmap-Status: unchanged Constraint: test-only typing fix; no behavior change Tested: mypy teaagent/ tests/ --explicit-package-bases (0 errors on the two files); pytest tests/test_real_usage_tui.py tests/test_real_usage_agents.py (145 passed, 6 skipped) Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…6 review, T5, hypotheses) and the ADR-0043 packet Owner decisions taken in session on 2026-10-09 and recorded, no status moved: - ADR-0031: extend once to no later than 2026-11-15, bound to a booked B1 owner-observed coding session (date to be booked); revert if it does not occur, without further review. The 2026-09-29 close had passed with no recorded disposition. New record docs/work-log/adr-0031-2026-10-09-extension-decision.md; ADR-0031 expiry + decision log; readiness checklist (item 0, new date); roadmap-status H4 Next Gate and owner-decisions paragraph; backlog row; held-spec index decision queue; execution-plan header. - ADR-0043: packet prepared, deletion deferred to 2026-12-09 — docs/plans/adr-0043-disposition-packet-2026-10-09.md (per-surface importer sweep incl. runtime import-time reachability from teaagent.cli for federated_sync/consensus/jit_approval_server, LOC and exclusive test counts 4,920 + 3,429, guards, recovery anchors, execution order, falsifiers). The review's Phase 3 numbers are corrected where the packet differs. - DR-006: three-month falsifier review recorded (F1 exit 0 on both bases, F2–F4 unexercised); T5 Q3 survey recorded as attempted, blocked (vendor docs unreachable from the recording environment), deferred ≤ 2026-12-31. - Friction log: four owner-validated hypotheses moved to `rejected` with the owner's 2026-10-09 validation note (a hypothesis entry may only be rejected or promoted; none became owner evidence); intake status restated; three older hypotheses already promoted to owner evidence F2/F3/F7 left open for the owner. - Review record §7.3; INDEX rows; G5 corpus figures restated (670 / 669 / 8 / 369 / 292); generated inventory, aging dashboard, and knowledge bundles regenerated. Evidence bundle rebuilt on the clean tree in the next commit. Action: G-P2-21 Roadmap-Status: unchanged Constraint: documentation only; records owner decisions made in session; no mode, gate, or horizon status moved Tested: bash scripts/verify_docs.sh (all checks pass except the evidence bundle, rebuilt next per the bundle protocol); pytest tests/test_docs_consistency.py tests/acceptance/test_docs_acceptance_count_accuracy.py (56 passed) Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…decision records)
Regenerate docs/generated/release-docs-evidence.{md,json} on the clean tree
and refresh the docs inventory hash for the regenerated bundle. The recorded
gate run is green (validate_docs_consistency.py in full test-quality mode,
report_docs_aging.py --check).
Action: G-P2-21
Roadmap-Status: unchanged
Constraint: generated artifacts only; no hand edits
Tested: python3 scripts/build_release_docs_evidence_bundle.py on a clean tree with dev extras installed (gate pass); bash scripts/verify_docs.sh after commit
Confidence: high
Agent: Claude Code
Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
The aging dashboard records per-file last-commit timestamps; committing the 2026-10-09 decision records changed them, so the generated dashboard and the inventory hash went stale immediately after the commit. Regenerated on the clean tree. The evidence bundle is rebuilt in the next commit because the previous bundle recorded the gate as failing on this stale dashboard. Action: G-P2-21 Roadmap-Status: unchanged Constraint: generated artifacts only Tested: report_docs_aging.py --check and generate_docs_inventory.py --check pass after regeneration Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…decision records, gate pass)
Regenerate docs/generated/release-docs-evidence.{md,json} on the clean tree
and refresh the docs inventory hash for the regenerated bundle. The recorded
gate run is green (validate_docs_consistency.py in full test-quality mode,
report_docs_aging.py --check).
Action: G-P2-21
Roadmap-Status: unchanged
Constraint: generated artifacts only; no hand edits
Tested: python3 scripts/build_release_docs_evidence_bundle.py on a clean tree with dev extras installed (gate pass); bash scripts/verify_docs.sh after commit
Confidence: high
Agent: Claude Code
Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…g tests PR #76 CI `lint` failed on `check_test_assertion_regression.py` against the merge base: `test_real_delivery_ledger_passes` dropped from 2 to 1 assertion when `assert ledgers` became a skip, and `test_shell_blocks_dangerous_commands` read as DELETED because its safe replacement had a new name. - tests/test_evidence_ledger.py: a second real assertion is restored after the skip guard — every discovered ledger is a file named `evidence-ledger.yaml` — before the validator loop. - tests/test_security_fixes.py: the governed-boundary test keeps the original name `test_shell_blocks_dangerous_commands` (same intent, many more assertions, never spawns a process), so the gate sees a strengthened test. No `Allow-test-weakening` trailer is used; nothing was weakened. Action: G-P2-21 Roadmap-Status: unchanged Constraint: test-only; assertion counts now >= base for both tests Tested: python3 scripts/check_test_assertion_regression.py --base origin/main (no weakened or deleted assertions); pytest tests/test_evidence_ledger.py tests/test_security_fixes.py (28 passed, 1 skipped, exit 0); ruff check + format --check Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
…commit, not the merge commit On `pull_request` events `HEAD` is GitHub's synthetic "Merge <head> into <base>" commit, which carries no trailers, so the review-institution job's "Lore trailers + claim-commit gate" step failed on PR #76 (and would fail on every PR) while passing on pushes to main. Both steps that read the HEAD message now resolve `COMMIT="${{ github.event.pull_request.head.sha || github.sha }}"` and inspect that commit's message. Push-to-main behavior is unchanged (`github.sha` is the pushed head there). Action: G-P2-21 Roadmap-Status: unchanged Constraint: CI workflow change only; same scripts, same hook entrypoint; no gate weakened Tested: YAML parses; python3 .githooks/commit-msg on this commit's own message passes locally; the failing step's log shows the merge-commit cause (job 113818109121) Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
johnteee
pushed a commit
that referenced
this pull request
Oct 9, 2026
…ssions, blocked v0.1.1 tag Adds §7.4 to the 2026-10-09 roadmap review: the ten pull requests that landed Phase 0, the release commit, and the eight dependency bumps (head, merge commit, and main CI run for each), the method used to reproduce the Dependabot bumps on current main with uv (floors byte-identical to the Dependabot diffs, lock regenerated, never hand-edited), the single red run on #81 and its root cause (fixed-sleep heartbeat liveness test, now deadline-based), the v0.1.1 tag that is annotated locally at 609657e but cannot be pushed from the recording environment (GitHub proxy returns HTTP 403 for tag pushes and the refs API) with the exact owner commands, and the items left for the owner (tag, 8.4 GB zero-filled /dev/sda file in the recording container, three older friction hypotheses still open). CHANGELOG gains Unreleased entries for the post-0.1.1 bumps and the test fix; roadmap-status gains a one-line landing note. Docs inventory regenerated. The release docs evidence bundle is rebuilt in the following commit on a clean tree (two-commit protocol), so verify_docs.sh passes only after that commit. Action: G-P2-21 Roadmap-Status: unchanged Constraint: documentation only; no horizon, milestone, goal, or gate status moved; no code change Tested: scripts/verify_docs.sh passes every check except the release docs evidence bundle, which the next commit rebuilds Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
Merged
11 of 13 tasks
johnteee
added a commit
that referenced
this pull request
Oct 9, 2026
…ndabot supersessions, blocked v0.1.1 tag All CI checks green on f1c6240 (CI run 37979721894); main run 881 green on the base. Documentation and generated artifacts only: §7.4 landing record, CHANGELOG Unreleased entries, roadmap-status landing note, rebuilt docs inventory / release docs evidence bundle / OKF current bundle. Action: G-P2-21 Roadmap-Status: unchanged Constraint: merge commit only; documentation and generated artifacts; no code change Tested: PR CI run 37979721894 green on all 21 non-skipped checks; scripts/verify_docs.sh passes on the head Confidence: high Agent: Claude Code Agent-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
mainhas failed CI on every run since 2026-09-29 (runs 851–858; jobstest (ubuntu-latest, 3.12)andacceptance-all). This PR records a whole-project roadmap review (docs/reviews/roadmap-review-and-improvement-plan-2026-10-09.md), lands its Phase 0 repairs, and records the owner decisions taken on 2026-10-09. Root causes of the 9 failing tests were reproduced off-runner: 6 need full git history (shallow checkout), 1 read gitignored.teaagent/delivery/, 1 assumedgit initcreatesmain, 1 ranrg(absent onubuntu-latest) and exposed a receipt gap where a failed command rendered silently. Fixes:fetch-depth: 0on the two jobs, skips-with-reason for environment-dependent tests, a scratch root for TUI/CLI scenario tests (they wrote into the repo's own.teaagent/), DNS-independent readiness tests via an opt-in fixture,[failed: …]/[outcome unknown]receipt rendering plus known-gap derivation (Gate: governance-gap), atyping.Anycount ratchet, job timeouts, an archive-growth signal on the aging dashboard, and the replacement of a test that executedrm -rf /,mkfs, anddd if=/dev/zero of=/dev/sdafor real (it now asserts the governance boundary and fails if any subprocess is spawned). Owner decisions recorded: ADR-0031 extended to ≤ 2026-11-15 bound to a booked B1 session (revert otherwise); ADR-0043 packet prepared, deletion deferred to 2026-12-09; DR-006 falsifier review and T5 outcome; four friction-log hypotheses dispositioned.Action ID
Risk Class
teaagent/change:run_evidence.py/run_receipt.pyreceipt rendering and known-gap derivation; no runner, approval, audit-emission, or permission-mode change;scripts/check_high_risk_paths.py --base origin/mainexit 0)Self-Review Checklist
General (all PRs)
ruff check+ruff format --checkpass (teaagent/, tests/, scripts/)mypy teaagent/ tests/reports 0 issuespython3 scripts/run_test_tier.py --tier smokepasses (203 passed, 1 skipped)main78.97%; this PR adds tests and removes nonecheck_root_module_count.py≤ 184 (177)check_complexity.py≤ 99 (60)verify_docs.sh)High-Risk (if risk class = high)
Documentation
Why / What / How / Done
Why. Red
mainblocks releases, Dependabot, and every claim the gates make trustworthy; the repo's own debt register misdiagnosed the cause as runner-specific.What. 25 commits: review record, CI config, test-environment hardening (shallow clone, default branch, root user, DNS,
ssh-keygen, SSL env, loopback, repo-root pollution), receipt honesty, ratchets, owner-decision records, ADR-0043 packet.How. Each failure re-run in a shallow clone, with full history, and against an
origin/mainworktree before classification; every repair verified with the test's real exit code; no assertion weakened; skips carryreason=.Done. Full suite on the final tree in this environment: 2 failed / 6,760 passed / 34 skipped before the last env fixes, both remaining failures container SSL env and now fixed; CI on this PR is the authoritative check.
🤖 Generated with Claude Code
https://claude.ai/code/session_0132tPtPaCjZgUpU2Ezgmrta
Generated by Claude Code