Mo Self Review

Compatibility: designed for Codex CLI and compatible agents. Requires a git working tree with a diff vs the base branch and the unhappy-path-audit skill (for Q1-Q3 depth).

Developer authority and blocker calibration

These rules apply to Codex, Claude, and every other compatible reviewing agent.

• A developer's explicit, recorded decision is authoritative for the reviewed change: accept it, capture its rationale in the decision log, and do not re-raise it as a finding. Revisit it only when the developer asks, or when materially new evidence proves that the stated assumptions no longer hold; present that evidence as a question, not as an override.
• Blocking is exceptional. Mark a finding blocking only when it has a concrete, demonstrated path to a merge-bar failure: correctness breakage, data loss/corruption, security exposure, hang/deadlock, resource leak with material impact, incompatibility, or required validation that is genuinely missing. State the failing inputs/state and consequence.
• A plausible concern, preference, maintainability improvement, micro-nit, unproven risk, or item already accepted by the developer is non-blocking (or omitted). Do not turn every observation into a blocker: if everything is blocking, the label has lost its meaning.
• When uncertain, classify non-blocking and ask for the developer's decision; never use a blocking label merely to force attention or another review round.

Resource Map

Running this skill IS the review (don't retype the long prompt)

Invoking this skill — /mo-self-review [target] or the Skill tool — runs the whole gate for you. First resolve the target from the args:

Resolve an explicit user base first, then a GitHub PR base for PR review, then a fresh/verified remote-tracking default. Never silently substitute a stale local main. If network access is unavailable, report the remote ref's object ID and known freshness limitation. Always report the base ref, resolved commit, and merge-base; do not guess main for a release-branch change.

Then execute (do not shortcut):

1. If a callable code-review workflow exists, launch it at high on the resolved target and state the base ("this branch compared to <base>"), appending: 多角度评审，第一性原则，系统性思考问题，涉及到的修改需要调研完整的功能闭环，unhappy path cover. Pass scope/skip instructions through. If that capability is absent, execute §1–§5 directly with the available repository and review tools; capability absence must not block the gate.
2. On results, apply §3 (trace each finding's functional closure to its terminal node; personally spot-check any cluster of refutations — a verifier can repeat one wrong call) and §5 (developer decisions are authoritative; severity LAST, calibrated to the merge bar; decision-log every won't-fix/known-gap with its reason; no finding without a concrete failure).
3. Present a converged, ranked findings list, each with a fix-or-decision-log recommendation — not another review round.

§1–§8 below are the methodology this executes; consult them when applying the discipline or running the gate manually. When this skill is surfaced only as background reference (not explicitly invoked), treat §1–§8 as guidance — do not auto-launch a workflow.

Purpose — break the review → modify loop

The review→modify→review loop repeats because each review pass is incremental: a new pass finds something the last one didn't (a fresh angle, a missed branch), and re-flags items already decided "won't fix." This gate front-loads one exhaustive, calibrated self-review of your own diff so the eventual PR review (human or bot) has nothing new to add → the loop ends.

Run it on your working diff BEFORE git push / opening / updating a PR. It is a gate: it either passes (§7) or produces a fix/decision list — not an endless stream of nitpicks.

Enforcement gate

Scope = committed changes vs merge-base, staged changes, unstaged changes, and untracked files in scope—not just the last file touched. Inspect git status --short and git ls-files --others --exclude-standard; also verify that any required generated/delivery artifact is not silently excluded by ignore rules.

1. Multi-angle (多角度) — run each lens as a separate pass

Do not spot-check. Sweep the whole diff once per lens; a defect invisible to one lens is obvious to another.

2. First principles (第一性原则)

Prove it breaks — don't report "looks like it might." For each candidate defect, state a concrete failure: inputs/state → wrong output / crash / hang. If you can't, exhaust every bypass path (defer, cancel watcher, timer, retry, guard) before ruling — then drop it. This is the single biggest source of review-loop noise: unproven "might" findings that get fixed, re-reviewed, and spawn more "might" findings.

3. Complete functional closure (功能闭环)

For every change, trace the entire loop it participates in, not just the edited line — most missed-in-review defects live one hop away, in the other half of the closure. Trace to the terminal node.

Rule: if you changed one arc of a closure, open and read the arcs that consume or reverse it (the reader for a writer, the restore for a backup, the Reset for a Call). A change is not reviewed until its closure is closed.

4. Unhappy-path coverage

Run the unhappy-path-audit skill's Q1–Q3 over the resources/waits/growth the diff touches:

• Q1 leak/double cleanup — every creation reaches one effective destruction owner (incl. transfers and error paths).
• Q2 hung — every explicit or implicit wait dependency has a guaranteed release; fail-fast/control paths must not queue behind the work they stop.
• Q3 OOM — every accumulation has a bound / recycle.

Apply its 5-gate false-positive filter (G1 full-graph, G2 can-fail/block, G3 bound/release, G4 line-reread, G5 calibrate-last) before keeping any finding.

5. Convergence discipline — this is what actually breaks the loop

1. Calibrate to the merge bar. Blocking means a concrete, demonstrated merge-bar failure—not merely a concern. Flag only real defects (correctness, data loss, material leak, hang, incompatibility, security) or genuinely required missing validation; state the failure path. Style / micro-nits: fix silently or skip — never loop on them. When unsure, use non-blocking. (Assign severity LAST, per unhappy-path-audit G5.)
2. Keep a decision log and respect it. Record every intentional design choice and every "won't fix / acceptable" item (with the why). A developer's explicit decision closes that item for this review: do not re-raise or relabel it blocking unless materially new evidence invalidates its stated assumption. Re-reviews and PR reviewers re-surface these constantly; a written decision lets you dismiss them in one line instead of re-litigating.
3. Verify before flagging. No finding survives without a concrete failure (§2) that passed the 5 gates (§4).
4. One thorough pass beats many incremental. The whole point: exhaust §1–§4 now so the next reviewer finds nothing. If you're tempted to "just fix this one and re-run," you're back in the loop — finish the sweep first.

6. How to run

On your own working diff (the default — this is a self gate):

• If a callable review workflow exists, use it for parallel discovery and then apply §3 closure + §5 discipline yourself.
• Otherwise walk §1 lens-by-lens → §3 closure → §4 Q1–Q3 → §5 directly. This is the complete supported fallback, not a degraded or blocked review.

For concurrency/lifecycle changes, build the invariant, transition table, ownership graph, wait-for graph, and generation boundary from references/concurrency-lifecycle.md. Derive the test matrix from semantic axes; do not reuse a remembered case list.

Proportional Go unit-test validation and race stress

For ordinary sequential changes, run the exact focused test(s) and the owning package once in normal mode with GOWORK=off, -mod=readonly, -count=1, and a bounded -timeout. Apply the adaptive race protocol below only when G-RACE-STRESS is triggered:

1. Build a minimal focused set from each newly added or modified TestXxx plus the individual existing regression test(s) that directly prove the changed behavior or transition. When a shared helper, package/global state, or background worker changes, choose the representative tests for the affected contract; the package-wide run in step 5 covers the broader interaction. If an issue, CI failure, or review comment names a failing TestXxx, that exact test is mandatory in the focused set; adjacent tests are not a substitute.
2. Prove the selection is non-empty: first enumerate it with GOWORK=off go test -mod=readonly -list, or verify that the test output names every intended test. A successful command whose -run expression matched nothing is not evidence.
3. Measure each exact test once under -race, excluding first-build time, and choose an adaptive repetition count. Read duration T from the test's terminal event emitted by GOWORK=off go test -mod=readonly -json, not the rounded package summary. With stress budget B and measured test duration T, use N = clamp(floor(B/T), 1, 100). Default B to 30 seconds; if T is absent, non-positive, or below timer resolution, use the upper cap N = 100. Adjust B for the change's risk and CI budget, and record T, B, and N. If a pre-fix reproduction has a known occurrence window, override the formula so the post-fix run covers that window; record why.
4. Run each focused test separately so a slow test does not reduce repetitions for a fast one: GOWORK=off go test -mod=readonly -race -count=N -timeout 120s -run '^TestA$' ./pkg/path. Independent commands may run in parallel when they do not contend for the same external resource. Keep repetitions of one test in the same process so leaked package/global state remains observable.
5. Then run the entire owning package once under the race detector: GOWORK=off go test -mod=readonly -race -count=1 -timeout 240s ./pkg/path.
6. If the package directly or transitively uses CGo, replace go test in all commands with .agents/skills/mo-dev/scripts/mo-cgo-test; follow the mo-dev environment setup. Do not silently skip tests because the local linker or runtime environment is incomplete.

Every repeated-stress command must contain an exact -run expression naming one individual test. Never apply adaptive -count=N stress to a package pattern or the repository; full-package race coverage is step 5 and runs only once.

Use a bounded, test-appropriate -timeout when needed. Normal tests, non-race -count=N, coverage runs, or one focused race run do not substitute for this gate.

The only routine exception is a measurement-only allocation/performance test whose oracle is invalidated by race-runtime bookkeeping. Isolate only that measurement behind //go:build !race, keep an equivalent functional test in the race build, and stress the functional test with the adaptive race budget. Never hide functional behavior or an ordinary timing assertion behind !race. For any other platform, build-tag, or test-kind constraint, report the exact test and technical reason; the gate remains blocked until the constraint is resolved or the reviewer explicitly accepts equivalent validation.

Before accepting the stress result, audit the test design against recurring MO flake classes:

• synchronize phases with channels, callbacks, barriers, or observable conditions; do not use time.Sleep or a tiny deadline as the scheduler;
• assert durable behavior, not a transient map entry, worker ownership, or which goroutine happened to make progress;
• register cleanup immediately so it runs after failed assertions too; restore package/global state and stop goroutines, timers, sockets, allocators, and other caller-owned resources;
• make topology, ordering, IDs, and map-derived choices deterministic; repeated -count=N runs share one test process and must not inherit prior-run state;
• use a generous outer deadline only as a hang guard unless timeout behavior is itself the contract under test.

Race success does not prove a timing-, allocation-, or instrumentation-sensitive oracle under non-race or coverage execution. Run the matching CI mode as additional evidence when the changed test depends on one of those properties. All evidence must contain the real exit status and be newer than the final semantic edit or rebase.

On a PR, use the same methodology against the verified PR base. A callable review workflow may accelerate discovery, but it does not replace personal closure and severity verification.

Depth delegation: for the leak/hung/OOM analysis, drive the unhappy-path-audit skill; for CGo build/test env and MO operator/format specifics, see mo-dev.

7. Exit gate — the diff is self-review-clean when ALL hold

□ every §1 lens swept over the whole diff
□ every changed arc's functional closure (§3) traced to its terminal node
□ Q1–Q3 unhappy paths (§4) checked on touched resources/waits/growth
□ state ownership, wait-for dependencies, and generation transitions modeled where applicable
□ every finding either FIXED or written to the decision log (§5.2)
□ severity calibrated to the merge bar (§5.1) — zero open blockers
□ every new/modified and directly affected Go behavioral unit test passed a proven non-empty focused run
□ when G-RACE-STRESS applies, focused adaptive -race -count=N and one owning-package -race run passed, with T/B/N recorded
□ every !race measurement-only test retains a race-tested functional counterpart when race validation applies
□ test matrix covers every changed transition and evidence is newer than the final edit/rebase
□ applicable domain guards passed (index-plugin → §8) — additive to the §1–§4 sweep above, never a substitute for it

Only then push / open the PR. If a subsequent PR review still finds a real blocker, that's a gap in §1/§3 coverage — add the missed lens/closure arc here so the gate catches it next time (the gate improves; the loop still ends).

8. Domain guard — index-plugin changes

A domain guard is a supplement to §1–§5, never a replacement. Always run the full multi-angle sweep over the ENTIRE diff (§1–§4) regardless of whether this guard applies; §8 only adds algo-specific checks when index-plugin files are touched. Passing §8 alone is not a review.

Apply when the diff touches index-algorithm dispatch, any pkg/vectorindex/<algo>/plugin/, pkg/fulltext/plugin, or pkg/indexplugin. Run mo-dev index-plugin reference §9 as the single source of current algorithm-specific checks. Its greps are candidate discovery only: inspect matched production code, exclude comments/tests, and prove registration, hooks, applicable ISCP/CDC wiring, build-tag behavior, and public-path coverage before passing or blocking. Do not copy the detailed guard here; duplicated rules drift.