[wikki26]

Scope

Focused follow-up for the outstanding deterministic guard/publish issues in #70. Target branch: harness/autopilot-2026-10-30; this does not modify that branch directly.

Fixes

• Parse changed paths with Git's NUL-delimited --name-status -z; use the exact path as an argument when reading a file's diff, rather than parsing Git-quoted patch headers.
• Scan every commit that the publisher's original bundle would transfer, including forbidden tier C/D path edits and synthetic secret-like content added and later removed. Reject agent merge commits.
• Promote edits that remove lines from existing tests out of tier A; such edits cannot qualify for automatic tier-A merge.
• Make --max-tier an actual failure after all tier promotions, including dependency manifests.

Regression evidence

Added 7 focused negative/positive tests to the existing tests/test_harness.py: tab-containing protected names, A-to-B rejection, valid A/A and B/B ceilings, removing an assertion without removing the test function, synthetic secret add/remove, protected path add/remove, and an untracked skip subsequently corrected in commit history. Adjusted the prior two-commit skip test expectation to reject unsafe published history.

• Hosted harness tests and workflow quality checks pass on this exact branch.
• Alex/Mohammad confirm guard restrictions match the intended policy.
• Integrate into harness: autonomous improvement loop and decisions.yaml lifecycle (design, off until founder approval on #43) #70, then rerun evaluation/activation gates. No activation or merge is authorized by this patch.

These changes are in CI guard tooling and tests only; no robot-control or safety runtime changes, workflow activation, secrets, or published agent branches.

Hosted evidence (31 October 2026)

• Run 41203133657: dedicated quality/test passed with 369/369 executed and zero skipped, including all seven added guard regression tests. ROS verifier smoke and governance baseline also passed.
• DCO passed on the feature branch. The generic reusable repository-quality / quality/test is optional/skipped here; the repository-local dedicated quality/test is the executed gate.
• The original harness: autonomous improvement loop and decisions.yaml lifecycle (design, off until founder approval on #43) #70 remains a draft. This patch is for Alex and Mohammad to inspect and incorporate, not permission to activate the harness or auto-merge robot changes.

[panthera-momagdii]

[P1] Apply the selected task's tier ceiling to every commit in the history the publisher transfers.

Reviewed #76 at head 6726eb294077709d4454171c657d70e93f749c2b against its parent #70 at 54bdd2b, including the full two-file patch, tools/harness_guard.py, harness_lib.py, policy tiers and prior #70 review findings. The NUL-delimited path parsing, existing-test tier promotion and add-then-remove secret/protected-path checks address the earlier examples. However, the new history loop checks C/D paths, secrets, skips and binaries, but does not reject a tier-B intermediate edit for a --max-tier A task. The new ceiling check at the end compares only the net-diff tier.

Reproduction (isolated Git history, no real secrets): from a base containing ros2/src/pkg/pkg/node.py, make signed commit 1 changing its return value (tier B); make signed commit 2 reverting the source file to its base contents and adding ros2/src/pkg/test/test_new.py containing def test_new(): assert True (tier A). The net base..HEAD diff contains only the new test (tier A), while git diff --name-status parent..commit on each transferred commit contains the tier-B source path. Under the current run() checks, --max-tier A is therefore not rejected; harness_publish.push_bundle() still publishes both commits. This violates the selected task's history/scope ceiling even though the final tree is tier A.

Please compute an effective tier ceiling over all published commits (including per-commit test-line removals/dependency-tier promotions), fail when any commit exceeds --max-tier, and add a two-commit source-edit/revert + tier-A-test regression. Preserve the strict C/D and secret-history rejection, human gates and draft-only rollout. This PR is stacked on #70; #70's separately reported stage-4 all-skipped/neutral verification gap also remains unresolved and is not cleared by this review.

Verification: I ran the two-commit history scenario in an isolated local Git repo and confirmed the net diff is test-only while both intermediate diffs modify node.py. I inspected, but did not execute, harness_guard.py itself. Hosted Repository quality run 41203133657, exact head: 369/369 checker tests passed, 0 skipped; this negative case is not covered. NOT RUN: actual harness publish, live dry run, workflow activation or robot tests. This review does not approve #70 or authorize harness activation.

[BotshareAI]

Thanks @wikki26 and @panthera-momagdii. Carried into #70 together with the follow-up 008633c, which answers the P1 here: the guard now computes a tier for every commit the publisher transfers, with the same promotions as the net diff (allowed tier D mission paths and dependency manifests to B, new tests in tier C areas to A, removed lines of existing tests to B); the effective tier is the highest of the net change and every commit, and --max-tier fails on it. Regressions in tests/test_harness.py: source edit then revert plus a new test (net diff test-only, rejected at --max-tier A, accepted at B), and a test weakened then restored. Both fail on 6726eb2. Please re-review on #70.

---

Generated by Claude Code