Orchestrata Ultimate › Plans › briefs › qwen › Q1-report

Q1 report — One C/C++ toolchain declaration: finish, test and verify

Source: plans/briefs/qwen/Q1-report.md at 8a2cb42ad · last changed 2026-10-11 · first committed 2026-10-11 · 245 lines.

Prose pending.

Plan text

Q1 report — One C/C++ toolchain declaration: finish, test and verify

Executor: Qwen3.8-max (openrouter/qwen/qwen3.8-max-0902) via opencode run. Worktree: /home/mohamed/workspace/wt-BARE-PYTHON, branch fix/bare-interpreter-gate (HEAD c1eebd019). Deliverable state: complete, tested, STAGED, not committed. No STOP condition occurred.

Done-when checklist (§7)

• [x] §6.1 passes, 0 skipped — 83 passed, 1 warning in 69.78s (first pass, before the two out-of-scope fixes; every module re-verified inside the final §6.6 run).
• [x] §6.2 outputs exactly as the spec expects (both pasted below).
• [x] ruff/black clean on every changed Python file; bash -n clean on deploy/toolchain.sh, deploy/install-orchestrata.sh, deploy/install-tools.sh.
• [x] Only the allowed files staged (§4 + the two orchestrator-accepted out-of-scope edits); generated files left unstaged. gate.py review-staged → commit review: PASS, 6 review(s), no blocking findings, review rc=0.
• [x] §6.6 resolver: "escalate": false, "untested": []. Bounded run through .run_exclusive.sh --force --json: rc 0; totals quoted from .q1/run-q1.json: "test_totals": {"passed": 1678, "failed": 0, "errors": 0, "skipped": 0}, "passed": 89, "failed": 0, "total": 89, "reused": {"modules": 0, "tests": 0} (--force: every selected test executed), time_seconds: 170.7, head_commit: c1eebd0196a76938ea2acd863d3a940309ce581b, dirty_sig: dad0683fb091a8a6844f053fc364876b95e2d69f.

Files changed (staged) and why

Pre-written by the orchestrator, verified against §3, staged unchanged unless noted:

Written/fixed by me per §4:

Out-of-scope edits (not in §4's table; both ACCEPTED by the orchestrator mid-task):

Left unstaged per §4/instructions: reports/run-tests-latest.json (runner-written, appeared modified), .q1/ (run scratch), .run_exclusive.sh (orchestrator's wrapper), plans/briefs/qwen/ (this report). The resolver also regenerated plans/reference-edges.json.affected-scratch (gitignored scratch — absent from git status). No other generated file (docs/CODEBASE_MAP*.md, plans/ast-*.json, context.json) appeared modified at any point.

Tests I had to fix in the pre-written modules, and which side was wrong

1. tests/scripts/test_worktrees_provision.py::test_cmd_provision_pins_toolchain_when_guix_gcc_shadows_system — the pre-written rewrite was wrong twice: (a) it deleted the helper _guix_shim_bin while this test still called it (guaranteed NameError — the module had never been run); (b) it left the test asserting the removed pre-§3 output (toolchain override: line from the PATH-shadow probe). Per §2 rule 5 I restored _guix_shim_bin verbatim, and per §6.1's rule ("a test asserting something §3 does not say is wrong") I re-pointed the assertions at the §3 output: toolchain: CC=/usr/bin/gcc CXX=/usr/bin/g++ (from <repo>/env/toolchain.env) plus dropped C_INCLUDE_PATH and provenance OK, with the Guix shim still first on PATH to prove PATH no longer decides. The shipped code was correct; only the test was fixed. Also updated the module docstring's stale "Guix-shadowed-gcc pin" phrase to the declaration wording.
2. tests/deploy/test_toolchain_sh.py — written but never run: passed as written, no fixes.
3. tests/deploy/test_install_tools.py — harness incomplete for the §3 design (the tests were right, the harness predated the installer's new sibling dependency); fixed as described above. No assertion changed.

§6 outputs

§6.1 (exact command, from the worktree root):

83 passed, 1 warning in 69.78s (0:01:09)

0 skipped; collection check confirmed all 10 new/rewritten deploy tests ran (incl. test_sync_stage_refuses_an_unusable_declaration[missing|bare-name|not-executable]). Post-fix re-verifications: tests/deploy/test_install_tools.py → 9 passed ... in 82.42s; tests/infra/test_gen_env_profiles.py tests/scripts/test_check_env_fleet_agreement.py → 32 passed ... in 11.16s; toolchain/envrc/provision subset → 41 passed in 10.37s.

§6.2 (real behaviour, outside pytest):

$ bash -c 'source deploy/toolchain.sh && load_toolchain env/toolchain.env && echo "CC=$CC CXX=$CXX CIP=${C_INCLUDE_PATH-unset}"'
toolchain: CC=/usr/bin/gcc CXX=/usr/bin/g++ (from env/toolchain.env)
CC=/usr/bin/gcc CXX=/usr/bin/g++ CIP=unset

$ env C_INCLUDE_PATH=/tmp/leak CC=gcc .venv/bin/python -c '...pin_system_toolchain(dict(os.environ))...'
toolchain: CC=/usr/bin/gcc CXX=/usr/bin/g++ (from /home/mohamed/workspace/wt-BARE-PYTHON/env/toolchain.env), dropped C_INCLUDE_PATH
/usr/bin/gcc /usr/bin/g++ False

§6.3: ruff --fix on the changed files → Found 1 error (1 fixed, 0 remaining) (the unused import shutil in scripts/worktree_toolchain.py); black → 9 files left unchanged (second pass on test_install_tools.py: ruff All checks passed!, black 1 file left unchanged); bash -n deploy/toolchain.sh deploy/install-orchestrata.sh deploy/install-tools.sh → clean.

§6.4 git status --porcelain after staging (final; .q1-affected.json already removed):

M  .envrc
M  deploy/install-orchestrata.sh
M  deploy/install-tools.sh
A  deploy/toolchain.sh
A  env/toolchain.env
M  plans/affected-tests-manifest.json
 M reports/run-tests-latest.json
M  scripts/build_ft_tree_sitter.py
M  scripts/build_janus_ft.py
M  scripts/hooks/pre-push
M  scripts/worktree_toolchain.py
M  scripts/worktrees.py
A  tests/deploy/test_envrc_toolchain.py
M  tests/deploy/test_install_orchestrata_cc.py
M  tests/deploy/test_install_orchestrata_toolchain.py
M  tests/deploy/test_install_tools.py
A  tests/deploy/test_toolchain_sh.py
M  tests/scripts/test_worktrees_provision.py
?? .q1/
?? .run_exclusive.sh
?? plans/briefs/qwen/

§6.5 (run twice — after the first staging and again after the two accepted out-of-scope edits; final):

commit review: PASS
  6 review(s), no blocking findings
review rc=0

§6.6 resolver (final staged set): escalate: False, untested: [], 71 paths → 89 modules. Reasons included: env/toolchain.env: keys CC, CXX changed — resolved to 52 test(s) that name them (RUL-032: no escalation); .envrc: resolved via 1 reference edge(s) (tests/deploy/test_envrc_toolchain.py); deploy/toolchain.sh: resolved via manifest-declared coverage.

§6.6 run — first attempt (spec's direct run_tests.py invocation) ended rc 1 with 1672 passed, 6 failed, 0 errors, 0 skipped · 86/89 modules: 3 failures in tests/deploy/test_install_tools.py (harness gap — fixed above, proven by clone/deploy/install-tools.sh: line 210: .../toolchain.sh: No such file or directory), 2 in tests/infra/test_gen_env_profiles.py + 1 in tests/scripts/test_check_env_fleet_agreement.py (tofu/PATH environment — see finding below; not caused by this change: both modules pass untouched once tofu is reachable). The authoritative rerun went through the host lock wrapper exactly as instructed:

$ mkdir -p .q1
$ bash .run_exclusive.sh --force --json <the 71 resolved paths> > .q1/run-q1.json 2> .q1/run-q1.err; echo "run rc=$?"
run rc=0

Totals from .q1/run-q1.json: "test_totals": {"passed": 1678, "failed": 0, "errors": 0, "skipped": 0}; "passed": 89, "failed": 0, "total": 89; "rerun": {"count": 0, ...}; "solo_retry": null; "reused": {"modules": 0, "tests": 0} (--force); time_seconds: 170.7. Stderr tail: [run_tests] ✅ All safety gates pass, [run_tests] 🧪 89 modules · 8 workers × 6 batches/worker · 420s timeout, [run_exclusive] 21:07:46 run_tests.py rc=0. All safety gates (source provenance, parallel safety 0.1–0.6, test purity, prolog boundary, rulings lint) passed; failed_modules: None.

Findings against the spec itself

1. §5.1's "Assert: rc 0 for the bootstrap stage reached" is unachievable as an installer-exit-code assertion. With the sanctioned recording-uv harness the installer always exits non-zero after the sync stage: the recording uv never creates a real .venv, so the free-threaded check fails (FAIL: deployed venv is not free-threaded) — documented as expected in _write_recording_uv's own docstring, and the pre-existing tests never asserted rc there. Implemented per the parenthetical: record_file.exists() is the proof the bootstrap (sync) stage was reached and uv sync exited 0; the installer's rc is not asserted in that test. The refusal tests do assert returncode != 0 (the loader exits before sync).
2. The spec's §6.6 wait loop self-matches. timeout 3600 bash -c 'while pgrep -f "scripts/run_tests.py" ...' — the bash -c wrapper's own command line contains the pattern, pgrep -f matches it (pgrep excludes only itself, not its ancestors), so the loop never exits and returns rc 124 → a false STOP on every invocation. I used the bracket form pgrep -f "scripts/run_tests[.]py" (same semantics, no self-match) until the orchestrator replaced waiting with .run_exclusive.sh. The pgrep approach also cannot guarantee non-overlap (a run can start between the last probe and the exec), which the lock wrapper fixes properly.
3. §6.6's assumption of an out-of-the-box "untested": [] was wrong for deploy/toolchain.sh (stem spelling + runtime-computed source — see the manifest entry above). The spec names the test file test_toolchain_sh.py in two places, so the manifest rule is the only fix consistent with §3.
4. PATH/tofu finding (environment, not this change). The first bounded run's 3 infra failures trace to infra/gen_env.py::tofu_binary(): $TOFU unset, tofu absent from the opencode session's PATH (it lives at /home/mohamed/bin/tofu, declared in pyproject external-tools: "installed standalone into ~/bin on this host"), so it fell back to terraform 1.5.7, which cannot init the required_version >= 1.6.0 root (Error: Unsupported Terraform Core version). Proof it is environmental: both modules pass untouched with ~/bin on PATH (32 passed in 11.16s). The orchestrator then put /home/mohamed/bin on the session PATH permanently.
5. Minor: scripts/worktree_toolchain.py's module docstring still opens with the 2026-09-23 shadow-pinning narrative ("Pinning CC/CXX to the system compiler"); the function docstrings (pre-written) carry the accurate declaration wording. Left untouched (§3: verify, do not redesign; §4 lists other files for word fixes). Removal/reword candidate for Mohamed.

Helpers that became unused (removal candidates — kept per §2 rule 5, only Mohamed authorises removals)

• _path_without_gcc(fakebin) in tests/deploy/test_install_orchestrata_cc.py — its only consumer was test_sync_stage_fails_closed_without_a_compiler_on_path, replaced per §5.1 (a PATH without gcc is no longer a meaningful refusal case: the declaration, not PATH, decides).
• _guix_shim_bin(tmp_path) in tests/scripts/test_worktrees_provision.py — deleted by the pre-written diff while still referenced; I restored it and it is used again by test_cmd_provision_pins_toolchain_when_guix_gcc_shadows_system, so it is NOT unused now; listed here only to document the delete/restore round trip.

Rule compliance notes

• No bare python/pytest/pip; every command ran via env -u PYTHONPATH .venv/bin/python ... from the worktree root. No new dependencies; pyproject.toml/uv.lock untouched. No mocks/stubs/fakes/monkeypatch/skips added; function-scope fixtures and tmp_path throughout; the recording-uv subprocess seam kept as is (§2 rule 3). Absolute imports only; the banned hedge word (§2 rule 4) appears in no added code, comment or docstring — the only hits in the staged diff are DELETED lines of the old tests. Nothing deleted: every pre-existing helper survives. No commit, push, merge, rebase, reset, stash or branch switch; nothing outside the worktree touched; 10.53.1.x never contacted; the full suite was never run — both run_tests.py invocations used the explicit resolved paths only.

Related

• Plans
• Orchestrata Ultimate — project index
• Workflow
• Tasks