@rosetears/aili-pi 0.1.0

package/skills/local-review-gate/SKILL.md

@@ -0,0 +1,207 @@
+---
+name: local-review-gate
+description: Run the explicit `/local-review` command or an equivalent user-requested standalone local audit over a diff, ref, PR, or OpenSpec change; do not trigger automatically from BUILD/SHIP, implementation completion, merge wording, or ordinary review-adjacent work.
+---
+
+# Local Review Gate
+
+## Purpose
+
+Use this skill when the user invokes `/local-review` or explicitly requests the same standalone report-first local audit before repair or PR submission.
+
+This skill is the workflow authority for the AILI-owned local review gate. It is not OpenCode's built-in `/review`, it does not replace `/ship`, and neither its report nor a subagent PASS is lifecycle acceptance or SHIP, release, merge, archive, or closeout evidence.
+
+## Non-Negotiable Boundaries
+
+- Do not override, replace, or depend on OpenCode's built-in `/review` command.
+- Do not claim release-ready, archive-ready, merge-ready, or SHIP-ready status; `/local-review` verdicts apply only to the reviewed target and evidence scope.
+- Do not use a `/local-review` result as lifecycle acceptance or SHIP/release/merge/archive evidence.
+- Do not push, create PRs, comment on GitHub, approve/request changes, merge, tag, publish, delete, reset, clean, or mutate remote state.
+- PR mode may run only the exact GitHub CLI allowlist `gh pr view`, `gh pr diff`, and `gh pr list --head`; do not run `gh api`, `gh pr checkout`, `gh pr comment`, `gh pr review`, `gh pr merge`, `gh pr create`, `gh repo clone`, or equivalent push, merge, comment, review, checkout, clone, or remote-mutating commands.
+- Do not perform repair until after a categorized report exists and `--repair` or an explicit user/current-contract approval authorizes repair.
+- Any delegated review context remains read-only and terminal. After an explicit `--repair`/current-contract authorization, ROSE may repair directly or use one independently justified fresh edit assignment; one affected recheck is sufficient and no old context is resumed.
+- Do not hide skipped checks, unavailable tools, stale evidence, partial context, or `Unverified` items.
+
+## Upstream Provenance and Adaptation
+
+Before copying or adapting upstream review material, read `references/upstream-provenance.md` and preserve its active/deferred source split. Active behavior references for this gate are:
+
+- `references/ecc-code-review-adaptation.md`: local/PR target selection, full-file review, validation, report artifact, verdict mapping, and read-only PR-mode boundaries.
+- `references/review-repair-lane-adaptation.md`: ECC-derived review/repair lane triggers, build-fix loop discipline, and separate repair ownership.
+- `references/orchestration-adaptation.md`: bounded capability selection, blocking/advisory split, evidence reconciliation, and main-agent-only integration/verdict rules; upstream fan-out language is non-authoritative where it conflicts with trigger-gated routing.
+- `references/addyosmani-code-review-rubric.md`: five-axis review rubric, `Critical`/`Important`/`Suggestion`, spec/task-first reading, concrete fixes, and uncertainty/proof gates.
+- `references/codex-github-compatibility.md`: behavior-only Codex/GitHub compatibility for AGENTS rules, PR focus instructions, high-priority findings, and local review/fix parity without cloud mutation.
+- `references/graphify-local-review.md`: existing-output-only adapter for optional architecture context from the official global Graphify skill. This gate neither owns nor launches Graphify and adds no install, registration, project execution, hook/plugin, scheduler, lifecycle gate, or completion authority.
+
+- Active derivative-reference sources: primary ECC `affaan-m/ECC` and `addyosmani/agent-skills`.
+- Behavior-only guidance: official Codex review docs; do not copy Codex documentation text.
+- Deferred or non-primary candidates: Kodelyth ECC, Elaine mirror/candidate, and `cminn10/ecc2cursor` unless a later scoped provenance review activates them.
+- Preserve MIT license/provenance for copied/adapted ECC or addyosmani material, including source URL, source path, copied/adapted scope, and adaptation rationale.
+- Replace Claude/ECC-specific tool, path, command, permission, and runtime assumptions with OpenCode/AILI-safe equivalents before activation.
+- Keep long copied/adapted material in references where possible, and use precise triggers and near-miss boundaries to avoid over-triggering.
+- Do not activate `.claude` paths, Claude-only tools, ECC-only command names, `ccg-workflow`, Codex/Gemini runtime dependencies, public `multi-*` commands, or unsafe remote mutation defaults.
+
+## Target Resolver
+
+Resolve and report the target before reviewing or accepting findings. Direct ROSE inspection is the default; no review lane is required merely because this command was invoked.
+
+| Input | Required resolution |
+|---|---|
+| no args | staged, unstaged, and untracked local changes; report absent categories too |
+| `--base <branch>` | current branch, selected base, merge-base, diff range, and any local staged/unstaged/untracked overlay |
+| `--commit <sha>` | commit identity, parent/range, touched files, and unavailable commit metadata |
+| `--pr <url\|number>` | read-only PR metadata/diff only through the exact GitHub CLI allowlist `gh pr view`, `gh pr diff`, and `gh pr list --head` when available; no `gh api`, checkout, clone, comment, review, merge, create, push, or equivalent remote mutation |
+| `--change <id\|path>` or OpenSpec path | proposal, design, tasks, specs, interview, test-plan, context, progress, drift log or legacy implementation notes, review report, final diff, and verification evidence |
+| `--focus <text>` or adversarial wording | target plus focus lane emphasis such as security, tests, data loss, performance, spec drift, or adversarial review |
+| `--repair` | only after a categorized report exists; convert approved findings into scoped repair packages and require re-review |
+
+For OpenSpec `--change <id|path>` targets, resolve the conventional artifact paths directly instead of relying only on broad glob/search output. Local `openspec/` directories may be git-ignored or omitted from snapshot-style indexes, so the target resolver must attempt exact reads or file listings for `proposal.md`, `design.md`, `tasks.md`, `interview.md`, `test-plan.md`, `context.md`, `progress.txt`, `review-report.md`, `drift-log.md`, legacy `implementation-notes.html`, and `specs/**/spec.md` before marking the artifact absent. If a broad search finds no OpenSpec files but exact reads/listing show the change exists, treat the broad search as insufficient evidence rather than as absence.
+
+If target, base, PR identity, OpenSpec change id, report path, permission, or required evidence source is ambiguous and cannot be safely inferred, return `BLOCKED` with the exact missing decision.
+
+## Full-Context Evidence Policy
+
+Diff-only review is insufficient for high-confidence findings.
+
+Before reporting a definite critical, security, data-loss, compatibility, correctness, or test adequacy finding, require:
+
+- full inspection of changed files and relevant untracked files;
+- related tests, fixtures, config, docs, AGENTS files, manifests, schemas, or protocol docs that constrain behavior;
+- callers, consumers, peer patterns, or downstream output paths when the claim depends on integration behavior;
+- command or artifact evidence for verification claims, or an explicit skipped reason with risk.
+
+When full context cannot be inspected within scope, label the affected claim `Unverified`, downgrade confidence, and prevent a `PASS` verdict unless the item is immaterial or explicitly accepted as risk.
+
+## Report Artifact Contract
+
+For non-trivial targets, create or update a repository-local Markdown report before repair starts.
+
+- OpenSpec default: `openspec/changes/<change-id>/review-report.md`.
+- Non-OpenSpec target: ask once for a repository-local report path, or offer chat-only review only when the user explicitly accepts the persistence limitation as `Unverified`.
+- Do not store or print secrets, tokens, private keys, cookies, raw logs, full transcripts, full file dumps, or private data in reports; use redacted path:line/type evidence instead.
+
+The report must record:
+
+- target, base/ref, merge-base, commit, PR metadata, or change id;
+- timestamp and current worktree/dirty-path classification when relevant;
+- source artifacts reviewed and source artifacts missing;
+- files/artifacts inspected in full, sampled, or skipped;
+- review lanes and checks run, skipped, stale, or unavailable;
+- categorized findings with severity, category, source lane, evidence anchor, affected file/artifact, required action, repair owner, repairability, current status, and re-review requirement;
+- previous findings, resolved/deferred/accepted-risk/out-of-scope items, and re-review history;
+- final verdict and named residual `Unverified` items.
+
+All review/test/security/coverage/convergence lane results embedded or referenced by the report must use the exact canonical finding/result envelope in `aili-delivery-flow/references/protocols/subagent-result.md`. A worker proposes `fix`, `refute-with-counter-evidence`, `accept-named-risk`, or `Unverified-block`; ROSE owns final disposition. Do not reconcile by majority vote, lane count, or averaged confidence, and do not treat a worker verdict as final authority.
+
+### Conditional arbitration artifact
+
+- Create or update exactly `openspec/changes/<change-id>/review-arbitration.md` only when a finding is disputed, blocking, cross-session, or materially inconsistent.
+- Preserve stable finding ID, competing claims, evidence and counter-evidence, proposed dispositions, ROSE disposition and rationale, decision owner, status, required recheck, freshness, and residual `Unverified` items.
+- A credible material minority finding stays open until fixed/rechecked, refuted with counter-evidence, accepted as a named risk by the authorized owner, or blocked as `Unverified`.
+- Do not create the artifact for routine uncontested findings and do not pre-create an empty instance.
+
+## Verdict Vocabulary
+
+Use only these local-review verdicts:
+
+- `BLOCKED`: target, permissions, evidence, report placement, or required decision is missing.
+- `NEEDS_FIXES`: actionable blocking or important findings remain.
+- `NEEDS_REVIEW`: human/product decision or conflicting evidence is required.
+- `PASS_WITH_UNVERIFIED`: no blocking findings remain, but named skipped or unverifiable items remain.
+- `PASS`: no blocking findings for the audited target and evidence scope, with fresh evidence and no material `Unverified` items.
+- `REPAIRING`: approved repair packages are being executed by separate edit/repair or edit/test lanes.
+- `REREVIEW_REQUIRED`: fixes were made or prior findings remain unresolved and need fresh review.
+
+`PASS` never means release-ready or SHIP-ready.
+
+For large or harness-sensitive `/local-review --change <id|path>` targets, `NEEDS_FIXES` and `BLOCKED` block BUILD continuation. `PASS_WITH_UNVERIFIED` may continue only after the user accepts each named `Unverified` item; otherwise return `NEEDS_REVIEW` or `BLOCKED` with the missing acceptance.
+
+## Review Capability Selection
+
+Select one primary review question and run the proactive delegation scan. Add at most one auxiliary capability when the user requests it or a concrete evidence gap triggers it; direct review is the no-trigger/blocked fallback. Default concurrency is at most two but is not a hard cap. Larger bounded read-only fan-out requires independent non-overlapping contexts, concrete benefit, suitable owners, and an explicit join plan. Do not run a review/test/security/coverage/convergence set without trigger evidence.
+
+- target/context lane: target identity, diff/base/PR/change scope, dirty paths, and report placement;
+- `code-reviewer`: correctness, maintainability, architecture, performance, and context adequacy;
+- `test-engineer` or `test-coverage-reviewer`: verification story, test quality, and uncovered paths;
+- `security-auditor`: auth, permissions, secrets, shell/installers, network, dependencies, storage, or data-loss risk;
+- `pr-test-analyzer`: branch/PR-like diffs, changed-test review, CI-log interpretation, and focused command matrix;
+- `ai-regression-scout`: prompts, agents, skills, routing, fixtures, model/tool routing, or generated-output changes;
+- `silent-failure-reviewer`: stale evidence, skipped gates, misleading reports, swallowed errors, and false PASS risk;
+- `convergence-reviewer`: formal OpenSpec, multi-phase, harness-sensitive, or evidence-heavy work where source artifacts must be compared to tasks, diff, progress, review, and verification evidence;
+- browser/E2E lanes only when the changed surface and artifact placement justify them;
+- open-source/provenance lanes only when public/package/provenance exposure is in scope.
+
+Any selected auxiliary packet must include target identity, accepted scope, source artifacts, diff/files, evidence already run, artifact placement rules, forbidden remote mutation, skipped-capability constraints, and expected terminal status. It is read-only, single-use, non-delegating, and must not repair its own findings or be resumed.
+
+For A33, each lane names exactly one packet-declared repository/cwd, references one current WT-001 context, re-reads applicable target rules that may narrow but never broaden, and writes any user-visible report to the owning repository. It records inspected scope, freshness, skipped checks, and the shared-trust soft-boundary limitation; it performs no broad host scan and never duplicates or rebinds identity, keys, approvals, Git state, rule bodies, verification command/cwd, or containment facts.
+
+Apply the adapted upstream evidence gates only to the selected question:
+
+- inspect the relevant accepted artifacts and current implementation/test evidence before a definite finding;
+- require `Critical`/`Important` findings to include exact evidence, concrete failure mode, why existing guards do not catch it, and a concrete fix;
+- accept zero findings only when inspected scope, skipped checks, and confidence are recorded;
+- treat an empty/status-less auxiliary result as missing evidence, not as permission to resume or dispatch an automatic retry;
+- split reconciled findings into blocking and advisory buckets, with unverifiable high-risk findings blocking until checked or explicitly accepted.
+
+## Convergence and Phase Checkpoints
+
+For formal or multi-phase work, compare the accepted source artifacts, relevant savepoints, final diff/worktree state, and fresh claim-matched checks. If parallel work was actually used, reconcile its compact status, evidence, and blockers before continuing.
+
+The active lifecycle owner remains outside `/local-review`: BUILD completion uses direct changed-scope/task-link inspection and the smallest relevant check, with one auxiliary capability only for a concrete gap and at most one targeted repair/recheck. `Package 12` is historical terminology for `complete-aili-workflow-orchestration`, not a generic review stage. A local-review report neither triggers nor satisfies lifecycle completion.
+
+In the historical `complete-aili-workflow-orchestration` queue, Packages 1–11 record lightweight savepoints containing complete behavior scope, files changed, unresolved items, and next package. Generic queues use the same lightweight savepoint fields without inheriting that numbering. Optional commands or diff feedback are not package closure or independent quality gates.
+
+## Anti-Pseudo-Completion Checks
+
+Reject pseudo-completion when:
+
+- prior findings are unresolved or not explicitly accepted/out-of-scope;
+- verification output is stale, missing, partial, or not tied to the exact reviewed target;
+- a lane reports success without anchors, inspected scope, lane status, and skipped-check reasons;
+- tasks, specs, test-plan items, or accepted requirements lack implementation/evidence coverage;
+- dirty paths are unknown or unrelated to the claimed scope;
+- skipped checks lack a reason and risk;
+- the report claims to replace PR review without target/base metadata, inspected files, checks, findings, and re-review evidence.
+
+## Repair and Re-Review Loop
+
+1. Resolve target and write/update the categorized report.
+2. Classify findings before repair.
+3. If repair is authorized, select one scoped repair package; ROSE works directly unless one fresh assignment independently passes the delegation gate.
+4. Apply at most one targeted repair and rerun the smallest affected check once.
+5. Reconcile the changed finding directly. Use a new read-only context only when changed evidence creates a separately justified capability gap; never resume the original context or launch an automatic review cycle.
+6. Keep verdict `REREVIEW_REQUIRED` until the affected finding has fresh direct or independently justified evidence; report any remaining blocker instead of starting another round.
+
+## Output Contract
+
+```text
+LOCAL REVIEW STATUS: BLOCKED | NEEDS_FIXES | NEEDS_REVIEW | PASS_WITH_UNVERIFIED | PASS | REPAIRING | REREVIEW_REQUIRED
+CONFIDENCE: HIGH | MED | LOW | VERY LOW | UNKNOWN
+
+TARGET:
+- Mode:
+- Scope/base/ref/commit/PR/change:
+- Report path or chat-only waiver:
+
+SOURCE ARTIFACTS:
+- Reviewed:
+- Missing/stale/unavailable:
+
+LANES AND CHECKS:
+- Ran:
+- Skipped with reason/risk:
+- Re-review required:
+
+FINDINGS:
+- [critical|high|medium|low|suggestion|unverified] category - evidence - required action - owner - status - re-review requirement
+
+REPAIR PLAN:
+- Authorized: yes | no
+- Packages or N/A:
+
+UNVERIFIED / ACCEPTED RISK:
+- <item or N/A>
+
+NEXT:
+- <fix, re-review, user decision, or ship remains separate>
+```

package/skills/local-review-gate/references/addyosmani-code-review-rubric.md

@@ -0,0 +1,33 @@
+# addyosmani Code Review Rubric Adaptation
+
+## Provenance
+
+| Field | Value |
+|---|---|
+| Upstream source | `https://github.com/addyosmani/agent-skills` |
+| Upstream HEAD | `8c6530305396f341b5da7201cf1f7e390fdb863f` |
+| Source paths | `agents/code-reviewer.md`, `skills/code-review-and-quality/SKILL.md` |
+| Source blobs | `96cac1d79edca4a9231cbe6af50415b5e4d6cf42`, `5efda7afb5d0e4a5393c5a7da84e15b197f7b5b6` |
+| License | MIT License, Copyright 2025 Addy Osmani |
+| Copy/adapt scope | Adapted review dimensions, severity vocabulary, spec/task-first review order, verification story, and actionable-fix rules; upstream prompt is not copied wholesale. |
+| Rationale | The upstream rubric is concise and matches AILI's existing `code-reviewer` lane. |
+
+## OpenCode / AILI adaptation boundaries
+
+- Keep AILI's hidden `agents/code-reviewer.md` as a read-only OpenCode subagent.
+- Do not route through `/review`; OpenCode owns that built-in command.
+- Do not make code-reviewer spawn other personas.
+- Do not convert suggestions into blockers unless evidence shows accepted-scope risk.
+
+## Activated AILI behavior
+
+- Read the spec, task, issue, or OpenSpec source artifact before judging implementation.
+- Review tests first when tests exist; tests reveal intent and coverage.
+- Evaluate every change across five axes: correctness, readability, architecture, security, and performance.
+- Require a verification story: tests reviewed, build/type/lint status if provided, manual/static evidence, skipped checks, and remaining `Unverified` items.
+- Use severity vocabulary: `Critical`, `Important`, and `Suggestion` in review-lane output, then map to local-review severities and verdicts during ROSE reconciliation.
+- Every Critical/Important finding needs file:line evidence, a concrete failure mode, a concrete fix recommendation, and a proof note explaining why existing guards do not already prevent it.
+- Use uncertainty/proof gates: if evidence is incomplete, downgrade, mark `Unverified`, or recommend a focused follow-up instead of guessing.
+- Zero findings is valid when the review records inspected scope, evidence, skipped checks, and confidence.
+- Prefer structural remedies that remove complexity or duplicate logic rather than style-only commentary.
+- Do not block for subjective rewrites when the change improves the codebase, follows project conventions, and has no accepted-scope risk.

package/skills/local-review-gate/references/codex-github-compatibility.md

@@ -0,0 +1,26 @@
+# Codex and GitHub Review Compatibility
+
+## Provenance
+
+| Field | Value |
+|---|---|
+| Source type | Official OpenAI Codex documentation and GitHub PR review behavior, used as behavior-only guidance |
+| Copy/adapt scope | No official documentation prose is copied; this file summarizes compatibility behaviors for AILI. |
+| Rationale | `/local-review` should produce evidence and reports that remain compatible with external Codex/GitHub PR review flows without depending on cloud mutation. |
+
+## OpenCode / AILI adaptation boundaries
+
+- Do not copy official Codex or GitHub documentation text.
+- Do not require Codex Cloud, GitHub automatic review, PR comments, PR reviews, or push permissions for local review.
+- Do not run `gh api`, `gh pr checkout`, `gh pr comment`, `gh pr review`, `gh pr merge`, `gh pr create`, `gh repo clone`, or equivalent remote-mutating commands in `/local-review --pr`.
+- Do not treat GitHub PR comment style as the only report format; AILI keeps persistent local reports first.
+
+## Activated AILI behavior
+
+- Keep `AGENTS.md` and project-local rules review-friendly: local review lanes must read the nearest applicable rules when they constrain findings.
+- Support PR-style focus instructions through `/local-review --focus <text>` and report the focus in the target metadata.
+- In PR mode, gather metadata and diff only through the exact GitHub CLI allowlist `gh pr view`, `gh pr diff`, and `gh pr list --head` when available.
+- Map high-priority PR-style findings into AILI `Critical` or `Important` findings with concrete failure modes and fixes.
+- Preserve review/fix parity locally: review produces a categorized report first; approved repair creates scoped edit/test packages; affected findings require re-review before closure.
+- Keep GitHub mutation as a separate, explicit, non-default user action outside `/local-review`.
+- If Codex/GitHub metadata is unavailable, mark that lane skipped or `Unverified` with reason and risk instead of claiming PR parity.

package/skills/local-review-gate/references/ecc-code-review-adaptation.md

@@ -0,0 +1,40 @@
+# ECC Code Review Command Adaptation
+
+## Provenance
+
+| Field | Value |
+|---|---|
+| Upstream source | `https://github.com/affaan-m/ECC` |
+| Upstream HEAD | `49128b5763b7ac0b50acef35ac0bcca08d1576af` |
+| Source path | `commands/code-review.md` |
+| Source blob | `2382c59968231c205f9cf9cffa9013f480899a9c` |
+| License | MIT License, Copyright 2026 Affaan Mustafa |
+| Copy/adapt scope | Adapted workflow structure and checklist concepts; no wholesale command copy is active. |
+| Rationale | ECC's command contains the closest local/PR review target model for `/local-review`. |
+
+## OpenCode / AILI adaptation boundaries
+
+- Do not add an ECC-named top-level command.
+- Do not override OpenCode's built-in `/review`.
+- Do not activate `.claude` paths, Claude-only tool names, `gh api`, GitHub review/comment/merge commands, or ECC artifact locations.
+- Do not activate remote mutation defaults; `/local-review --pr` remains read-only.
+- Codex/GitHub official docs remain behavior-only references and are not copied here.
+
+## Activated AILI behavior
+
+- Select mode from explicit target syntax: default local changes, `--base <branch>`, `--commit <sha>`, `--pr <url|number>`, `--change <id|path>`, `--focus <text>`, or `--repair`.
+- Gather a target inventory before findings: changed files, untracked files, base/ref/commit/PR/change metadata, unavailable metadata, and skipped target lanes.
+- Read changed files in full before making high-confidence correctness, security, data-loss, compatibility, or test findings; this is the active full-file review rule.
+- Inspect relevant tests, config, docs, AGENTS rules, manifests, source artifacts, and callers/consumers when a finding depends on integration behavior.
+- Classify findings with severity, category, evidence anchor, required fix, owner, status, and re-review requirement.
+- Run available validation appropriate to the repository; skipped or unavailable checks must include reason and risk.
+- Write a persistent categorized report for non-trivial targets before repair begins.
+- Map ECC `APPROVE`/`REQUEST CHANGES`/`BLOCK` into local-review verdicts: `PASS`, `PASS_WITH_UNVERIFIED`, `NEEDS_FIXES`, `NEEDS_REVIEW`, `BLOCKED`, `REPAIRING`, and `REREVIEW_REQUIRED`.
+- Treat zero findings as valid only when target scope, full-context inspection, and verification/skipped-check evidence are recorded.
+
+## Rejected upstream behavior
+
+- Publishing GitHub reviews, approving PRs, requesting changes on GitHub, posting comments, or using GitHub API comment endpoints.
+- Falling back from PR mode to unsafe shell parsing or raw user-provided PR arguments.
+- Storing reports in `.claude` directories or any ECC-specific artifact tree.
+- Treating validation pass alone as review approval without source/context inspection.

package/skills/local-review-gate/references/graphify-local-review.md

@@ -0,0 +1,33 @@
+# Graphify Local-Review Adapter
+
+[FRAME] This adapter may consume already-produced output from the official globally installed Graphify skill as optional architecture context. It does not install, register, configure, launch, update, or remove Graphify.
+
+## Positive trigger
+
+- Use this adapter only inside an explicit `/local-review` when an existing Graphify result is already available and directly relevant to an architecture, ownership, boundary, coupling, or hotspot review question.
+- Consume at most one scoped result; do not ingest a complete graph or broad report when the review question can be answered from current source and tests.
+- If no usable result exists, continue without Graphify. Do not create or refresh one from this gate.
+
+## Evidence routing
+
+- Treat Graphify as upstream snapshot/navigation evidence only.
+- Use CodeGraph or current files for exact symbols, implementations, call paths, tests, dirty-state impact, and current behavior.
+- Reconcile every Graphify lead against current source, tests, accepted artifacts, and Git state before reporting a finding.
+- When Graphify conflicts with current evidence or its freshness is unknown, current evidence wins and the Graphify claim remains `Unverified`.
+- Do not duplicate the same discovery through Graphify, CodeGraph, and broad grep; choose one owner for the question and narrow follow-up reads.
+
+## Upstream ownership
+
+[KNOWN|EXTERNAL] The accepted official CLI and global Agent-Skills commands are `uv tool install graphifyy` and `graphify install --platform agents`. Source: immutable Graphify `v0.9.20` README and `graphify/install.py` recorded in `openspec/changes/improve-context-maps-and-session-handoffs/design.md`.
+
+- Installation and global registration are owned by the guided installer, not `/local-review`, and require separate exact approvals.
+- A future project build, update, query, path, or explain operation is a different operation with its own exact target/effect approval.
+- This repository owns no Graphify wheel, lock, profile, venv, project index, parser, launcher, hook, plugin, backend, server, scheduler, or support/qualification claim.
+
+## Local-review boundaries
+
+- Never invoke `graphify`, `/graphify`, `uv`, a Graphify backend, or a Graphify server from this gate.
+- Never install the OpenCode platform plugin or mutate `.opencode`, hooks, `AGENTS.md`, Git state, or project graph output.
+- Never turn Graphify output into correctness, security, acceptance, completion, SHIP, merge, archive, or release evidence.
+- Keep absolute paths, private data, full graph payloads, and raw logs out of the review report; cite a bounded existing artifact/result and current-file evidence instead.
+- Findings remain advisory until the normal local-review evidence and verdict rules resolve them.

package/skills/local-review-gate/references/orchestration-adaptation.md

@@ -0,0 +1,51 @@
+# Review Orchestration Adaptation
+
+## Provenance
+
+| Field | Value |
+|---|---|
+| Upstream source | `https://github.com/affaan-m/ECC` |
+| Upstream HEAD | `49128b5763b7ac0b50acef35ac0bcca08d1576af` |
+| Source paths | `commands/orch-review.md`, `commands/multi-plan.md`, `commands/multi-execute.md`, `commands/multi-backend.md`, `commands/multi-frontend.md`, `commands/multi-workflow.md` |
+| Source blobs | `5216c7df157a9099214c122bc096e9693730ac77`, `b50912b1f0900ed34935228d973dd92902b97342`, `167a9b559e54b2caa4ad0700635cf5c9b1a367d9`, `95cc95d831d65779d86203f446b043cbee0af66a`, `fc1c402d985f46baf31cacf465d1db2f6c0ce3da`, `5458945c22a20eaa43eb332f53db7a7da1d2a408` |
+| License | MIT License, Copyright 2026 Affaan Mustafa |
+| Copy/adapt scope | Adapted evidence reconciliation, fail-closed missing-evidence handling, dedupe, and adversarial finding checks; upstream fan-out and phase machinery are not active defaults. |
+| Rationale | Selected ECC evidence patterns strengthen proactive trigger-gated ROSE review without creating untriggered automatic lane dispatch or a broader command surface. |
+
+## Secondary provenance
+
+| Field | Value |
+|---|---|
+| Upstream source | `https://github.com/addyosmani/agent-skills` |
+| Upstream HEAD | `8c6530305396f341b5da7201cf1f7e390fdb863f` |
+| Source path | `references/orchestration-patterns.md` |
+| Source blob | `09cddd31d52a95be1d9bca814547d184b9945e44` |
+| License | MIT License, Copyright 2025 Addy Osmani |
+| Copy/adapt scope | Adapted direct invocation, single-persona command, optional independently justified evidence joins, and research isolation; automatic fan-out and Claude-only runtime details are not active. |
+
+## OpenCode / AILI adaptation boundaries
+
+- Do not add public `multi-*` commands, ECC-named commands, or router personas.
+- Do not activate external workflow runtimes, Agent Teams, Claude-specific team/tool assumptions, Codex/Gemini runtime dependencies, or `ccg-workflow`.
+- ROSE/MainAgent is the only orchestrator and the only final decision owner.
+- Personas do not invoke other personas; a worker may return one concrete missing-capability need to ROSE but cannot start a follow-up lane.
+- Review, security, convergence, and test-analysis lanes are read-only unless their specific lane contract is a bounded test runner with no source edits.
+
+## Activated AILI behavior
+
+- Direct invocation remains preferred for one perspective on one artifact.
+- Single-persona command behavior stays scoped to `/local-review`; it routes target resolution and report state, not a broad router command.
+- Direct ROSE review is the default. At most one auxiliary capability may use up to two fresh independent read-only contexts only when a concrete evidence gap and concurrency benefit justify them.
+- Each lane packet carries target identity, accepted scope, source artifacts, diff/files, verification evidence, skipped-lane constraints, artifact placement, forbidden remote mutation, expected status vocabulary, and missing-evidence handling.
+- Treat a selected auxiliary result that fails, is empty/status-less, lacks anchors, or cannot inspect required context as missing evidence; do not resume it or dispatch an automatic retry.
+- Dedupe findings by affected file/artifact, failure mode, and required action; preserve the highest defensible severity.
+- Split findings into blocking and advisory buckets: confirmed Critical/Important or unverifiable high-risk findings block; suggestions and refuted findings remain advisory.
+- Use adversarial verification for Critical/Important findings: check whether context, tests, types, existing guards, or project rules disprove the finding before blocking.
+- When independently justified parallel evidence was actually used, reconcile its compact results before relying on the joined claim; do not add phase checkpoints or lanes merely because work has phases.
+
+## Rejected upstream behavior
+
+- A workflow tool owning approval or publish decisions.
+- Raw PR argument shell use.
+- Any GitHub comment, review, approve, merge, checkout, clone, push, or API mutation from review mode.
+- Replacing `/ideate`, `/define`, `/build`, or `/ship` with ECC multi-agent commands.

package/skills/local-review-gate/references/review-repair-lane-adaptation.md

@@ -0,0 +1,46 @@
+# Review and Repair Lane Adaptation
+
+## Provenance
+
+| Field | Value |
+|---|---|
+| Upstream source | `https://github.com/affaan-m/ECC` |
+| Upstream HEAD | `49128b5763b7ac0b50acef35ac0bcca08d1576af` |
+| Source paths | `agents/code-reviewer.md`, `agents/security-reviewer.md`, `agents/pr-test-analyzer.md`, `agents/build-error-resolver.md`, `commands/build-fix.md`, and ECC catalog entries for E2E, documentation lookup, TDD, verification loop, API/backend/frontend/language/domain skills |
+| Source blobs explicitly inspected | `af791188ac87321f749a96f140a85c739303f453`, `568351bbc710009acba3e4c32282c436291a34c8` |
+| License | MIT License, Copyright 2026 Affaan Mustafa |
+| Copy/adapt scope | Adapted lane duties, repair-loop safety, build-fix loop, confidence gates, and trigger boundaries; no upstream agent is copied wholesale. |
+| Rationale | ECC supplies useful specialist review/repair lane seeds, but AILI keeps review read-only and routes fixes through separate edit/test lanes. |
+
+## OpenCode / AILI adaptation boundaries
+
+- Review lanes remain read-only and must not repair their own findings.
+- Repair work must be explicit, scoped, separately owned, verified, and re-reviewed.
+- Do not activate Claude-only tools, ECC tool names, `.claude` artifact paths, `ccg-workflow`, or Codex/Gemini runtime dependencies.
+- Do not add dependencies, lockfile changes, or remote mutation commands unless a separate user-approved task explicitly authorizes them.
+- Use existing AILI skills where equivalent skills already exist instead of importing duplicate ECC skill bodies.
+
+## Activated AILI behavior
+
+### Review lanes
+
+- `code-reviewer`: use confidence filtering, exact evidence anchors, proof requirements for Critical/Important findings, false-positive suppression, and zero-findings-is-valid behavior.
+- `security-auditor`: trigger on auth, permissions, secrets, shell/installers, dependency changes, network, storage, user input, and data-loss surfaces; report exploit path and concrete mitigation without editing.
+- `pr-test-analyzer`: map changed behavior to tests, quality of assertions, changed-test risk, coverage gaps, and focused verification commands.
+- `test-engineer` / coverage lanes: verify behavior-first tests, regressions, skipped checks, and adequacy of the verification story.
+- `ai-regression-scout` and `silent-failure-reviewer`: trigger on prompts, agents, skills, routing, static fixtures, false PASS risk, swallowed errors, stale evidence, or misleading reports.
+- Browser/E2E lanes: only when UI/browser behavior changed and artifact placement is approved; no durable screenshots/traces without repository-local placement.
+
+### Repair lanes
+
+- Build/type repair follows ECC `build-fix` discipline: detect the build system, group errors, fix one error class at a time, rerun focused verification, and stop on repeated failure, deeper architecture need, dependency need, or more-errors-than-fixed regression.
+- Documentation repair is allowed only when behavior/config/public workflow docs are made wrong by the task; otherwise report docs drift for a scoped follow-up.
+- Refactor cleanup is advisory unless the current task explicitly includes cleanup or the finding blocks acceptance.
+- E2E/browser repair requires explicit artifact placement and must avoid production data mutation.
+- Domain/language review is conditional: activate TypeScript/Node, Python, shell, docs, packaging, API/backend/frontend, database, ML/CUDA, or mobile lanes only when the changed surface uses that domain.
+
+## Deferred or rejected ECC lane candidates
+
+- Loop operators and harness optimizers are not activated as autonomous agents; AILI keeps orchestration under ROSE and uses bounded package queues.
+- Language/domain specialists are reference checklists, not new top-level commands or always-on personas.
+- Documentation lookup remains source-driven research only when current user/task contract allows external lookup and no secrets are exposed.

package/skills/local-review-gate/references/upstream-provenance.md

@@ -0,0 +1,61 @@
+# Upstream Review Provenance
+
+## Scope
+
+[COMPUTED from Package 5 upstream scout evidence] This reference records the active and deferred upstream candidates for adapting local-review behavior, without copying full upstream prompts, commands, agents, or documentation text in this package.
+
+[COMPUTED from Package 5A implementation evidence] Tasks 5A.2-5A.8 are now activated through focused AILI reference files, core reviewer updates, orchestration contract updates, and static fixture checks rather than public ECC command imports or wholesale upstream prompt dumps.
+
+## Active source selection
+
+[COMPUTED from Package 5 upstream scout evidence] `affaan-m/ECC` and `addyosmani/agent-skills` are the active derivative-reference sources for future `/local-review` adaptation work.
+
+[COMPUTED from Package 5 upstream scout evidence] Official Codex review documentation is behavior-only guidance for review/fix-loop compatibility and must not be copied as prompt or documentation text.
+
+[INFERRED from Package 5 upstream scout evidence] `sifxprime/kodelyth-ecc`, `elainesmithburkel-code/everything-claude-code`, and `cminn10/ecc2cursor` are not primary sources for this change and remain deferred or non-primary unless later provenance, trigger-fit, and OpenCode compatibility checks justify activation.
+
+## Candidate inventory
+
+| Source URL | License | Source path(s) | Confidence | Intended use | Status | Adaptation notes |
+|---|---|---|---|---|---|---|
+| `https://github.com/affaan-m/ECC` | MIT | `commands/code-review.md` at SHA `2382c599...`, size `8179` | HIGH from upstream scout packet | Primary review-command reference for target selection, full-file/context reading, validation, verdict/report, PR-mode lessons, and repair/re-review workflow ideas | active-reference | Preserve source URL/path/SHA and MIT notice for copied/adapted material; adapt into AILI `/local-review` references rather than adding an ECC-named command; replace Claude/ECC paths, tool calls, agent names, and runtime assumptions with OpenCode/AILI-safe equivalents; avoid broad prompt copying in this package. |
+| `https://github.com/affaan-m/ECC` | MIT | `commands/orch-review.md` at SHA `5216c7d...`; `commands/build-fix.md` at SHA `568351b...`; `commands/multi-plan.md` at SHA `b50912b...`; `commands/multi-execute.md` at SHA `167a9b5...`; `commands/multi-backend.md` at SHA `95cc95d...`; `commands/multi-frontend.md` at SHA `fc1c402d...`; `commands/multi-workflow.md` at SHA `5458945...` | HIGH from current task evidence and focused fetch | Evidence reconciliation, blocking/advisory split, adversarial finding checks, and bounded repair concepts; upstream fan-out/phase machinery is reference-only | active-reference | Activated only through proactive trigger-gated ROSE rules; public `multi-*` commands, untriggered automatic swarms, external workflow runtimes, and remote mutation behavior are rejected. |
+| `https://github.com/affaan-m/ECC` | MIT | `agents/code-reviewer.md` at SHA `af79118...`; `agents/security-reviewer.md`; `agents/pr-test-analyzer.md`; `agents/build-error-resolver.md`; related ECC catalog skills under `.agents/skills/*` | HIGH for explicit seeds; MED for catalog-wide categories | Review/repair lane seeds: confidence/proof gates, security, test coverage analysis, build-error repair, E2E, docs lookup, TDD, verification loop, API/backend/frontend/domain/language review checklists | active-reference with conditional triggers | Activated as lane/checklist behavior only; no wholesale agent copy, no reviewer self-repair, no dependency changes, and no `.claude`/Claude-only tool assumptions. |
+| `https://github.com/addyosmani/agent-skills` | MIT | `agents/code-reviewer.md` at SHA `96cac1d...`, size `3801` | HIGH from upstream scout packet | Active rubric reference for five-axis review, Critical/Important/Suggestion classification, spec/task-first reading, concrete fix guidance, and uncertainty discipline | active-reference | Preserve source URL/path/SHA and MIT notice for copied/adapted material; adapt the review discipline into local review lane rules rather than pasting the complete upstream agent; reconcile severity vocabulary with local-review verdicts and existing AILI report tables. |
+| `https://github.com/addyosmani/agent-skills` | MIT | `skills/code-review-and-quality/SKILL.md` at SHA `5efda7a...`; `references/orchestration-patterns.md` at SHA `09cddd3...` | HIGH from current task evidence and focused fetch | Five-axis review evidence, change sizing, verification story, severity, direct invocation, single-persona command, optional evidence joins, and research isolation | active-reference | Activated under proactive trigger-gated AILI review rules; untriggered automatic fan-out and Claude-specific platform sections are not active. |
+| Official OpenAI Codex review docs | Official docs; do not copy text | GitHub review docs covering `@codex review`, automatic PR review, nearest `AGENTS.md` review guidelines, focus instructions, and high-priority findings | HIGH for behavior guidance from upstream scout packet | Compatibility guidance for local PR-style review, focus text, high-priority finding mapping, and local review/fix parity | active-reference, behavior-only | Do not copy docs text; use only behavior patterns. Keep `/local-review --pr` read-only unless separate approval authorizes remote mutation outside review mode. |
+| `https://github.com/sifxprime/kodelyth-ecc` | MIT | `commands/code-review.md` at SHA `8189f951...` | MED from upstream scout packet | Separate or fork-like comparison candidate if primary ECC leaves ambiguity | deferred | Treat as non-primary. Use only for later comparison after provenance and divergence checks; do not mix with primary ECC-derived material without recording source-specific scope. |
+| `https://github.com/elainesmithburkel-code/everything-claude-code` | MIT per upstream scout packet | mirror/candidate paths not activated in this package | LOW from upstream scout packet | Possible mirror/candidate for broad Claude-code review patterns | deferred | Lower confidence; do not activate without fresh provenance and copied-scope evidence. |
+| `https://github.com/cminn10/ecc2cursor` | MIT | candidate paths not activated in this package | LOW from upstream scout packet | Possible cursor-port reference, not a primary review source | rejected for primary review source; deferred for compatibility comparison | Do not use as the authoritative review source for `/local-review`; revisit only if later Cursor/OpenCode compatibility comparison is explicitly scoped. |
+
+## Copy/adapt rules for future packages
+
+[COMPUTED from accepted Package 5 task contract] Preserve MIT license/provenance whenever ECC or addyosmani material is copied or adapted.
+
+[COMPUTED from accepted Package 5 task contract] Record source URL, source path, commit/SHA when known, copied scope, adapted scope, and adaptation rationale in the affected reference or implementation file.
+
+[COMPUTED from accepted Package 5 task contract] Replace Claude/ECC-specific tool names, file paths, command names, agent names, permissions, and runtime assumptions with OpenCode/AILI-safe equivalents before activating imported behavior.
+
+[COMPUTED from accepted Package 5 task contract] Avoid over-triggering by adding precise trigger and near-miss boundaries for every adapted agent, lane, skill, or reference.
+
+[COMPUTED from accepted Package 5 task contract] Keep long copied/adapted material in reference files rather than always-loaded prompts whenever possible.
+
+[COMPUTED from accepted Package 5 task contract] Do not copy Codex documentation text; use official Codex review docs only for behavior-compatible local workflow guidance.
+
+[COMPUTED from accepted Package 5 task contract] Do not copy/adapt ECC agents, ECC skills, orchestration commands, or addyosmani rubrics wholesale in this provenance package; later packages must perform scoped adaptation and verification separately.
+
+## Package 5A activated reference files
+
+| Task | Activated artifact(s) | Evidence status |
+|---|---|---|
+| 5A.2 | `references/ecc-code-review-adaptation.md`; `/local-review` skill pointers | active/completed after static checks |
+| 5A.3 | `references/review-repair-lane-adaptation.md`; `agents/code-reviewer.md`; review-pipeline lane notes | active/completed after static checks |
+| 5A.4 | `references/review-repair-lane-adaptation.md`; conditional ECC skill catalog mapping | active/completed after static checks |
+| 5A.5 | `references/orchestration-adaptation.md`; `review-pipeline`; `parallel-subagent-dispatch`; `subagent-task-packet.md` | active/completed after static checks |
+| 5A.6 | `references/addyosmani-code-review-rubric.md`; `agents/code-reviewer.md` | active/completed after static checks |
+| 5A.7 | `references/codex-github-compatibility.md`; `/local-review` skill and command boundaries | active/completed after static checks |
+| 5A.8 | `scripts/harness_fixture_check.py`; `docs/harness/fixtures/command-routing-fixtures.yaml` | active/completed after static checks |
+
+## Limits
+
+[UNVERIFIED] This reference is not legal advice and does not analyze license compatibility beyond recording the upstream license/provenance evidence. Re-check source URLs, license files, and exact copied scope before any future bulk copy/adaptation.