diff --git a/agent-ops/.version b/agent-ops/.version index 2053fc91..683ec8c6 100644 --- a/agent-ops/.version +++ b/agent-ops/.version @@ -1 +1 @@ -1.1.147 +1.1.148 diff --git a/agent-ops/skills/common/code-review/SKILL.md b/agent-ops/skills/common/code-review/SKILL.md index ceba23cb..9851b471 100644 --- a/agent-ops/skills/common/code-review/SKILL.md +++ b/agent-ops/skills/common/code-review/SKILL.md @@ -90,10 +90,10 @@ Milestone task group contract: Review routing rules: - Route reviews so a normal mixed backlog can use both lanes substantially. This is not a per-review quota; neither `local` nor `cloud` is the default catch-all. -- `local`: first-pass and review-detectable work, including multi-file or API/call-site reviews when scope and required context are bounded and verification can be rerun or reasoned about locally. +- `local`: review-detectable work where edited files, direct call sites, and rerunnable tests are enough to judge correctness. - `local-G01` through `local-G10` and `cloud-G01` through `cloud-G10` are valid. Do not treat high grade alone as an implicit cloud route; context volume is separate lane evidence. - Repair non-behavioral review artifact drift during review instead of failing solely for it when implementation correctness, tests, and contracts remain judgeable. -- `cloud`: reviews requiring stronger-than-local judgment, larger-than-local context, or evidence recovery: weak verification that review cannot compensate for, security/auth, storage/migration, concurrency, protocol/schema, broad cross-domain behavior with hard-to-review failure modes, repeated non-trivial Required issues, or untrusted verification evidence. +- `cloud`: reviews where a wrong but plausible implementation could pass local tests because correctness depends on behavior promised outside the edited files, wider context, repeated Required issues, or evidence trust. - `cloud-G07` or higher is mandatory only when terminal-agent or external interactive behavior is central and cannot be made deterministic for local review, or when the review requires more source/test/diff/log context than local can hold without omitting material evidence: shell/CLI workflow implementation, bin script orchestration, process control, stdout/stderr parsing, exit-status contracts, long-running command diagnosis, terminal benchmark-style tasks, real bin/smoke/integration failures after unit tests passed, TUI/PTY/browser/screen repaint/cursor stream behavior, rerun-mismatched verification trust failures, or very large cross-module/cross-domain changes that cannot be safely split for review. Directory states: @@ -236,13 +236,13 @@ If the user-review gate does not trigger, write new routed plan/review files usi - Header tag is `REVIEW_`. - If the selected task group is `m-`, write the follow-up plan/review under the same `m-` task group. - Base the follow-up scope directly on the archived review findings. Keep it narrow and actionable. -- Choose lane/grade again; preserve the prior route when it was adequate. Raise `local-GNN` for explicit, deterministic, review-detectable follow-ups, but move `local -> cloud` when the failure exposes reasoning, risk, evidence, or context-volume gaps that local review should not own. +- Choose lane/grade again; preserve the prior route only when still adequate. Raise `local-GNN` for explicit, deterministic, review-detectable follow-ups, but move `local -> cloud` when edited files, direct call sites, and tests were not enough. - If the finding is only missing or untrusted verification evidence that an appropriate follow-up lane can produce by rerunning commands or collecting artifacts, do not write `USER_REVIEW.md` solely for that reason. Write a follow-up plan that reruns the exact commands in an appropriate lane and records actual stdout/stderr. - If follow-up verification involves mobile/UI hangs, timeouts, or no progress for 2 minutes, stop blind retries; require focused rerun output plus screenshot/window/UI-tree evidence path, or the exact reason capture is impossible. - Write an `Archive Evidence Snapshot` section in every WARN/FAIL follow-up plan. Copy enough prior-loop evidence from the archived plan/review logs for the next implementer to act without rereading archive by default: current archived plan/review log paths, verdict, Required/Suggested/Nit summary, affected files, verification evidence, and any roadmap/spec carryover. If exact prior context is still required, cite the specific archive file paths allowed to read; do not instruct the implementer to search `agent-task/archive/**` broadly. - Copy the follow-up plan's `Archive Evidence Snapshot` section exactly into the follow-up review stub. - Before choosing the follow-up route, apply this escalation gate: - - If the archived plan was `local-*` and the verdict is `FAIL` for correctness, completeness, test coverage, or verification trust, keep the follow-up on `local` with a higher grade when the issue is explicit, deterministic, and review-detectable; move to `cloud` only when the failure shows local capability or evidence trust is likely insufficient. + - If the archived plan was `local-*` and the verdict is `FAIL` for correctness, completeness, test coverage, or verification trust, keep the follow-up on `local` only when the fix is visible from edited files, direct call sites, and rerunnable tests; otherwise move to `cloud`. - If the follow-up work is terminal-agent work that cannot be reduced to deterministic commands and evidence, use `cloud-G07` or higher. - If unit tests passed but a real bin/smoke/integration command failed and local follow-up cannot isolate the failure with deterministic evidence, use `cloud-G07` or higher. - If the task depends on interactive TUI/PTY/browser/external CLI behavior, screen repaint/cursor stream parsing, or live command-palette state that cannot be captured deterministically, use `cloud-G07` or higher. diff --git a/agent-ops/skills/common/plan/SKILL.md b/agent-ops/skills/common/plan/SKILL.md index 8c1a4b13..8ea65174 100644 --- a/agent-ops/skills/common/plan/SKILL.md +++ b/agent-ops/skills/common/plan/SKILL.md @@ -97,10 +97,10 @@ Routing rules: - Build routing should be calibrated so a normal mixed backlog can use both lanes substantially. This is not a per-task quota: choose from the task's evidence, but do not let either lane become the default catch-all. - `local-G01` through `local-G10` and `cloud-G01` through `cloud-G10` are valid. Do not treat `G06+` or high complexity alone as an implicit `cloud` threshold; context volume is separate lane evidence. -- Use a higher `local-GNN` when risk can be contained by explicit planning, focused code review, deterministic tests, split subtasks, and bounded required context. -- Use `cloud-GNN` when the task needs stronger reasoning, larger-than-local context, or evidence recovery than local should be expected to provide: unclear external behavior, weak or missing verification for user-visible behavior, high-risk architecture judgment, security/auth/storage/concurrency/protocol changes with hard-to-review failure modes, repeated local failure, verification trust failures, or required source/test/call-site/log/review context that local cannot hold without omitting material evidence. +- Use a higher `local-GNN` only when the edited files, direct call sites, and rerunnable tests are enough to judge correctness. +- Use `cloud-GNN` when correctness depends on behavior promised outside the edited files, especially when a plausible but wrong test could still pass. - Use `cloud-G07` or higher only when terminal-agent or external interactive behavior is central and cannot be made deterministic for local handling, or when the task requires more source/test/diff/log context than local can hold without omitting material evidence: shell/CLI workflow implementation, bin script orchestration, process control, stdout/stderr parsing, exit-status contracts, long-running command diagnosis, terminal benchmark-style tasks, TUI/PTY/browser/screen repaint/cursor stream behavior, real bin/smoke/integration failures after unit tests passed, or very large cross-module/cross-domain changes that cannot be safely split. -- Review uses the same balance: choose `local` for review-detectable work with bounded scope, bounded context, and rerunnable verification, and choose `cloud` for reviews requiring frontier-level judgment, larger-than-local context, weak tests, security/auth, storage/migration, concurrency, protocol/schema, cross-domain behavior, repeated Required issues, or untrusted verification evidence. +- Review uses the same balance: choose `local` when edited files, direct call sites, and tests are enough to judge correctness, and choose `cloud` when behavior promised outside edited files, wider context, repeated Required issues, or verification trust decide correctness. - Treat non-behavioral review artifact drift as reviewer-repairable unless it prevents judging implementation correctness, tests, or contracts. - Treat obvious non-behavioral source nits as reviewer-repairable too: typos, stale comments, docs, or formatting only, with no behavior/test/API contract change. - Raise `GNN` with scope, ambiguity, missing tests, irreversible behavior, and blast radius after lane selection; keep grade model-independent. @@ -194,15 +194,15 @@ Complete all items below before creating active plan/review files. Work through GXX is an output of analysis, not an input. Determine lane and grade only after Step 2 is fully complete. - [ ] **Assess change scope** — count affected files, interface impact, call-site count, and the amount of source/test/review context that must stay loaded together. -- [ ] **Check risk factors** — mark any that apply: concurrency, storage/migration, protocol/schema, auth, irreversible behavior. +- [ ] **Check risk factors** — mark any that apply: behavior promised outside edited files, persisted-state rule, permission boundary, shared-state risk, irreversible behavior. - [ ] **Evaluate test confidence** — judge whether existing tests sufficiently verify the changed behavior. - [ ] **Decide lane** — choose between `local` and `cloud` after weighing both sets of signals. Across a normal mixed backlog, both lanes should appear regularly; do not force a quota for one task, and do not let either lane become the catch-all. - - Prefer `local` when the task can be made explicit, deterministic, and review-detectable; when affected files, call sites, and required context are bounded; when tests or focused smoke checks can be rerun locally; or when split subtasks can contain risk. + - Prefer `local` when a wrong implementation would be visible from the edited files, direct call sites, and rerunnable tests. - Prefer `cloud` when the task is too large for local context to hold the required source files, tests, call-site graph, generated artifacts, logs, or review evidence without omitting material context. - Prefer `cloud` when split subtasks still require broad cross-module or cross-domain context to be understood safely, even if each edit could be applied locally. - Prefer `cloud` when required behavior depends on unclear external systems or user-visible behavior that cannot be verified deterministically. - Prefer `cloud` when verification is weak or missing in a way that review cannot reliably detect. - - Prefer `cloud` when security/auth, storage/migration, concurrency, or protocol/schema risk has hard-to-review failure modes. + - Prefer `cloud` when a wrong but plausible implementation could satisfy local tests while breaking behavior promised outside the edited files. - Prefer `cloud` when prior local attempts failed for non-trivial reasoning or evidence-trust reasons. - Prefer `cloud` when terminal-agent work is central: shell/CLI workflow implementation, bin script orchestration, process control, stdout/stderr parsing, exit-status contracts, long-running command diagnosis, or terminal benchmark-style tasks. - Prefer `cloud` when real bin/smoke/integration verification failed after unit tests passed.