Home / Projects / Orchestrata Ultimate / Decisions / ADR index / ADR-0105

ADR-0105: The components branch pushes only after the affected-test gate learns the packages/** layout — a 345-commit push is not shipped past the gate with --no-verify

← ADR-0104 · ADR-0106 →

The record

Status: Accepted (ruling, Mohamed, 2026-10-18, on the push-decision prompt, verbatim: "Remediate the gate, then push") · Date: 2026-10-18

Context

Session 271 (2026-10-18), Mohamed's instruction: "clean up and push" on the wt-components project. The cleanup completed in five commits (chore gov/out ignore cd59b9b97; VOLKS lift v11 landing 5d8265be5; SPEC-AUDIT reverify evidence 028993f1d; WIKI toolchain scripts 0822e643f; session context/docs sync 191765874) and the working tree is clean. The branch components (HEAD 191765874) is 345 commits ahead of origin/master, has never been pushed, and no origin/components exists.

The push is blocked, and every push path to it is blocked by the same finding:

• git push origin components runs the pre-push hook (scripts/hooks/pre-push), which runs uv run python scripts/run_tests.py --affected origin/master...HEAD. The affected resolver (scripts/affected_tests.py) reports 7,176 changed files as UNTESTED, and run_tests.py --affected returns 1 on any untested file (fail-closed, _UNTESTED_IS_NOT_A_FULL_SUITE). Git aborts the push (failed to push some refs).
• deploy/ship.sh (the sanctioned push path) measures its blast radius from the last verified tree (reports/.ship-verified → 8a413deed, 2026-10-14) and runs the SAME run_tests.py --affected <base> — 7,178 UNTESTED from that base. The run-evidence-reuse skip cannot fire because no green run exists at this source id. Both paths are blocked identically.

UNTESTED breakdown (range origin/master...HEAD), by subtree:

By extension: 6,505 .py, 568 .json, 22 .toml, 20 .mmd, and a tail of lock/err/feature/ xml/gitignore files.

Root cause: the affected-selection model predates the multi-package layout. Only top-level src/ and tests/ are recognized coverage homes. The component lift (ADR-0103, VK-01 scaffold) lands components under packages/<pkg>/** with parity suites under packages/<pkg>/tests (measured on disk: contracts 128, kernel 298, governance 164, knowledge_graph 120, session 78, volks 220 test files) that run in per-package venvs — each packages/<pkg>/pyproject.toml is its own uv project, and the root pyproject.toml declares none of them as dependencies, so the root runner cannot even collect those suites. The gate therefore cannot see any of the packages' verification: component-wave parity reports (volks v11: parity PASS 32/32), lander records, per-package test runs, import_each.py. No pushed commit on origin/master ever touched packages/** — there is no precedent to follow.

Decision (proposed)

The components branch pushes only after the affected-test gate learns the packages/** layout. A 345-commit push is NOT shipped past the gate with --no-verify. The gate extension is a bounded packet:

1. an affected_tests resolution source mapping packages/<pkg>/** to the package's own parity test tree (which genuinely covers it);
2. a runner seam so run_tests.py executes per-package suites in the package's own venv (or a root wrapper the runner invokes and verifies) — the parity suites are not collectable under the root venv today;
3. manifest rules for the machine-inert artifact classes: the .dsh/skills symlink tree → a link-tree test; .realize-out ledgers and design JSONs; packages/*/uv.lock → the lock parity check VK-01 already claims;
4. then verify via deploy/ship.sh and push. Until the packet is green, the branch stays unpushed; the five cleanup commits remain on it.

Alternatives rejected

• git push --no-verify now (bare bypass): ships 340+ commits without the gate ever passing on the range. The UNTESTED stop is fail-closed by design, and the ship.sh doctrine's PUSH_NO_VERIFY is sanctioned only because "this IS the hook's work, already done" — here the hook's work has not been done. Bypass would also leave the gate gap in place for every future packages/** commit.
• 6,505 root-level wrapper tests (one per UNTESTED file under tests/): padding tests that assert nothing real, against the ≥90%-coverage and real-assertions doctrine.
• Declaring packages/** doc-inert: a false claim — they are code with real suites.

Consequences

• origin/master and prod are untouched either way; the components branch simply stays local until the gate packet lands.
• The gate packet itself is verifiable without a push: running run_tests.py --affected after the extension proves the selection resolves, and the parity suites run in their own venvs.
• After the first green ship, ship.sh stamps the verified source id and subsequent pushes bound their range from it, per the recorded-run design.

Provenance

• Mohamed's instruction, session 271, 2026-10-18: "clean up and push"; ruling on the Proposed record the same day, verbatim: "Remediate the gate, then push".
• Gate output measured the same day: run_tests.py --affected returns 1 on 7,176 UNTESTED files (range origin/master...HEAD) and 7,178 (range from last verified tree 8a413deed).
• Mechanism citations: scripts/hooks/pre-push; scripts/affected_tests.py (_UNTESTED_IS_NOT_A_FULL_SUITE, manifest plans/affected-tests-manifest.json); scripts/run_tests.py §2b (untested → return 1); deploy/ship.sh (blast radius from the last verified tree, PUSH_NO_VERIFY rationale).

Related ADRs

• Also mentioned: ADR-0103

Rulings cited

None.

Specs cited

None (or no spec page exists for the ones named).