refine(oss): action-item change/severity fields; replace sentinel commit gate with in-context auth.

- Add `change` (code/test/docs/config/ci/style/refactor) and `severity` (1–5) to ACTION_ITEMS schema
- Update action item table format to include Change and Severity columns
- Route by `change` field when Codex unavailable (code/refactor→sw-engineer, test→qa-specialist, docs→doc-scribe, style→linting-expert)
- Add `--agent` override to bypass routing table unconditionally
- Update TaskCreate description to carry change+severity for agent context
- git-commit.md: replace Path A/B/C + sentinel files with 3-signal in-context model (in-message, in-workflow, in-confirmation); rename to Commit Authorization + Push Authorization; no filesystem state or touch/rm needed
- foundry/.claude-plugin/plugin.json: bump to 0.20.0
- oss/resolve/SKILL.md: fix Step 4 skip condition — checkout runs unconditionally in pr mode; Step 8 skip when SELECTED_ITEMS empty
- oss/.claude-plugin/plugin.json: bump to 0.9.4

---
Co-authored-by: Claude Code <noreply@anthropic.com>

Author: Borda

plugins/foundry/.claude-plugin/plugin.json

@@ -15,5 +15,5 @@
   "license": "MIT",
   "name": "foundry",
   "repository": "https://github.com/Borda/AI-Rig",
-  "version": "0.19.1"
+  "version": "0.20.0"
 }

plugins/foundry/rules/git-commit.md

@@ -108,64 +108,61 @@ Never skip trailers because skill template omits them.
 
 ## Branch Safety
 
-Default branch is repo-specific — do NOT hardcode `main` or `master`. Hook detects dynamically via `git symbolic-ref refs/remotes/origin/HEAD`, `gh repo view`, or `git remote show origin`. Committing to default branch requires **second sentinel** (Gate 2 below).
+Default branch is repo-specific — do NOT hardcode `main` or `master`. Detect dynamically via `git symbolic-ref refs/remotes/origin/HEAD`, `gh repo view`, or `git remote show origin`.
 
 Before any `git commit`, check current branch:
 
 ```bash
 CURRENT_BRANCH=$(git branch --show-current)
 ```
 
-On default branch: two sentinels required (Gate 1 + Gate 2). On feature branch: one sentinel required (Gate 1 only).
+Feature branch: commit proceeds when any authorization signal present (§Commit Authorization). Default branch: `AskUserQuestion` always required — no signal overrides this.
 
-## Commit Gate (two gates)
+## Commit Authorization
 
-**Gate 1 — commit authorization** (all branches):
+Authorization determined in-context at the moment of committing — no sentinel files, no filesystem state.
 
-Sentinel path: `/tmp/claude-commit-auth-<repo-slug>-<branch-slug>` · TTL: 15 min
+**Feature branch — authorized when any signal present**:
 
-**Gate 2 — default-branch protection** (default branch only):
-
-Sentinel path: `/tmp/claude-commit-default-<repo-slug>-<branch-slug>` · TTL: 5 min (must touch immediately before commit)
-
-Slug algorithm: all non-alphanumeric → `-`, lowercased, consecutive dashes squeezed, trailing dashes stripped.
+| Signal | What it looks like |
+| --- | --- |
+| **In-message** | Current user message contains unambiguous commit instruction: "commit", "commit this", "make a commit" |
+| **In-workflow** | Running skill that names commit as documented step AND user invoked that skill — skill invocation = implicit authorization |
+| **In-confirmation** | User confirmed via `AskUserQuestion` in the current response turn |
 
-```bash
-# Compute both sentinel paths
-REPO_SLUG=$(git rev-parse --show-toplevel | xargs basename | tr '[:upper:]' '[:lower:]' | tr -cs 'a-z0-9' '-' | tr -s '-' | sed 's/-$//')
-BRANCH_SLUG=$(git branch --show-current | tr '[:upper:]' '[:lower:]' | tr -cs 'a-z0-9' '-' | tr -s '-' | sed 's/-$//')
-SENTINEL="/tmp/claude-commit-auth-${REPO_SLUG}-${BRANCH_SLUG}"
-DEFAULT_SENTINEL="/tmp/claude-commit-default-${REPO_SLUG}-${BRANCH_SLUG}"
-```
+**Default branch — always AskUserQuestion**: none of the above signals override default-branch protection. Always invoke `AskUserQuestion` before committing to default branch.
 
-**Path A — skill pre-auth** (skills committing as part of workflow, e.g. `/oss:resolve`):
-- `touch $SENTINEL` at start of commit phase; `rm -f $SENTINEL` on finish or abort (use `trap` to guarantee cleanup)
-- If committing to default branch: also `touch $DEFAULT_SENTINEL`; `rm -f $DEFAULT_SENTINEL` in same `trap`
-- Gate 1 sentinel exists and <15 min old (Gate 2 valid if on default branch) → hook allows commit
-- Sentinel absent, expired, or branch mismatch → hook blocks → fall through to Path B
+**When no signal present (feature branch)**: invoke `AskUserQuestion` before `git commit`. Question must show: target branch, whether default branch, diff size (`N files, +A −B lines` from `git diff --stat HEAD`), draft commit message subject line.
 
-**Path B — user ad-hoc request**:
-- No Gate 1 sentinel → invoke `AskUserQuestion` before `git commit`
-- Question must show: target branch, whether default branch, diff size (`N files, +A −B lines` from `git diff --stat HEAD`), draft commit message subject line
-- On user confirmation:
-  - Feature branch: `touch $SENTINEL` → `git commit` → `rm -f $SENTINEL`
-  - Default branch: `touch $SENTINEL && touch $DEFAULT_SENTINEL` → `git commit` → `rm -f $SENTINEL $DEFAULT_SENTINEL`
-- Never self-create sentinels without `AskUserQuestion` first — bypasses Gate 2 entirely
+**Never commit autonomously**: authorization must come from one of the three signals — never self-authorize without user signal in the current turn.
 
 ## Staging and Hooks
 
 - Never `git add -A` or `git add .` — always stage specific files by name
 - Never `--no-verify` — if pre-commit blocks, fix underlying issue
 - Never `--no-gpg-sign` unless user explicitly requests it
 
-## Push Safety
+## Push Authorization
+
+Same signal model as §Commit Authorization — no AskUserQuestion when signal present. Note: `git push` is not pre-allowed in settings; harness will prompt once regardless (by design).
+
+**Feature branch — authorized when any signal present**:
+
+| Signal | What it looks like |
+| --- | --- |
+| **In-message** | Current user message contains unambiguous push instruction: "push", "push this", "push the branch" |
+| **In-workflow** | Running skill that names push as documented step AND user invoked that skill |
+| **In-confirmation** | User confirmed via `AskUserQuestion` in the current response turn |
+
+**When no signal present**: invoke `AskUserQuestion` before `git push`. Show: target branch, remote, whether default branch.
+
+**Authorization scoped**: "commit this" does not authorize "push this" — push requires its own signal in the current message or turn.
+
+**Default branch**: `AskUserQuestion` always required — no signal overrides this.
 
-- **Never push without explicit user confirmation** — always invoke `AskUserQuestion` before any `git push`, including branch pushes, PR pushes, and release tags (prose question alone is not sufficient)
-- Authorization scoped: "commit this" does not authorize "push this"; ask separately for every push
-- Applies inside skill workflows — if skill (e.g. `/resolve`) includes push step, treat as "propose and confirm", not "auto-execute"; stop after committing, report what ready to push, wait for user to say push
-- Never push in autonomous bug fixing or as "final step" without being explicitly asked in that message
-- Never force-push (`--force`, `--force-with-lease`) to main/master — forbidden even with explicit user instruction
-- Never force-push to any other branch without explicit user instruction; prefer regular push with explicit confirmation
+**Force-push**:
+- Main/master: forbidden even with explicit user instruction
+- Other branches: only with explicit user instruction in current message; never as autonomous "final step"
 
 ## History Safety
 

plugins/oss/.claude-plugin/plugin.json

@@ -15,5 +15,5 @@
   "license": "MIT",
   "name": "oss",
   "repository": "https://github.com/Borda/AI-Rig",
-  "version": "0.9.2"
+  "version": "0.9.3"
 }

plugins/oss/skills/resolve/SKILL.md

@@ -260,7 +260,10 @@ Read every comment, review, inline code comment. Per inline code comment: if its
 | `[info]` | Praise, acknowledgement, emoji-only — skip |
 | `[self-review]` | Finding from `/oss:review` report — not a GitHub commenter; author = agent name |
 
-Build `ACTION_ITEMS`: `[{id, type, author, summary, file, line, url, full_comment_text}]` — `url`: `html_url` from GitHub API response; blank for report items
+Build `ACTION_ITEMS`: `[{id, type, change, severity, author, summary, file, line, url, full_comment_text}]` — `url`: `html_url` from GitHub API response; blank for report items
+
+- **`change`** — type of change required: `code` · `test` · `docs` · `config` · `ci` · `style` · `refactor`; infer from finding text and affected file path; default `code` when ambiguous
+- **`severity`** — estimated impact severity 1–5 (5 = highest): 5 = blocks merge / data loss / security; 4 = user-visible regression; 3 = correctness/important quality gap; 2 = improvement, non-blocking; 1 = nit / style; map from comment urgency and `type` (`[req]` → floor 3)
 
 ### Sources confirmation
 
@@ -270,17 +273,19 @@ Print action item table — **MUST render as markdown table; never use key-value
 
 - **Summary**: ≤60 chars — truncate at word boundary, append `…`
 - **Notes**: ≤45 chars — truncate; full text preserved in `full_comment_text`; use `—` when empty
+- **Change**: one of `code` · `test` · `docs` · `config` · `ci` · `style` · `refactor`
+- **Severity**: integer 1–5; `[req]` floor = 3
 
 Status codes: `pending` · `✓ resolved` · `⊘ skipped` · `⊘ no action`. Verbose reason → Notes column:
 
 ```markdown
 ### Action Items — PR #<number>
 
-| # | Type | Author | Status | Summary | Notes |
-|---|------|--------|--------|---------|-------|
-| 1 | [gh][req] | @reviewer | pending | rename param `x` to `count` | — |
-| 2 | [gh][suggest] | @maintainer | pending | add docstring | — |
-| 3 | [gh][question] | @reviewer | pending | why not use X instead? | — |
+| # | Type | Change | Severity | Author | Status | Summary | Notes |
+|---|------|--------|----------|--------|--------|---------|-------|
+| 1 | [gh][req] | code | 4 | @reviewer | pending | rename param `x` to `count` | — |
+| 2 | [gh][suggest] | docs | 2 | @maintainer | pending | add docstring | — |
+| 3 | [gh][question] | code | 3 | @reviewer | pending | why not use X instead? | — |
 ```
 
 Long content never justifies switching to key-value or separator-delimited format — truncate, stay in table.
@@ -335,7 +340,7 @@ Options: (a) Apply all [req] (X items) · (b) Apply all (N items) · (c) Skip al
 If per-item control needed: advise re-run after reducing source (e.g. use `report` mode instead of `pr + report`, or `--no-challenge` to cut upstream findings).
 
 Resolve `SELECTED_ITEMS`:
-- "Skip all" or no selections → `[]` → skip Steps 4–8, jump to Step 9
+- "Skip all" or no selections → `[]` → skip Step 8, jump to Step 9 (checkout + conflict resolution still run)
 - "Apply all [req]" → all `[req]` IDs
 - "Apply all" → all pending IDs
 - "Apply selected" → checked IDs from item questions
@@ -369,7 +374,7 @@ Create tasks **only for `SELECTED_ITEMS`** — not all pending items; avoids con
 ```text
 TaskCreate(
   subject="<type> <summary> — PR #<number>",   # <type> = full string with brackets
-  description="Author: @<author> | File: <file:line or '—'> | <full_comment_text>",
+  description="Author: @<author> | Change: <change> | Severity: <severity> | File: <file:line or '—'> | <full_comment_text>",
   activeForm="Implementing: <summary>"          # <summary> truncated to 80 chars
 )
 ```
@@ -378,7 +383,7 @@ Store returned task ID in each `SELECTED_ITEMS` entry as `task_id`.
 
 ## Step 4: Checkout PR branch
 
-*Runs only when `SELECTED_ITEMS` non-empty (set in Step 3e). Empty → skip to Step 9.*
+*Skip only when `MODE = report` with no PR# (`$PR_NUMBER` unset — no remote branch to check out). In pr mode, runs unconditionally regardless of `SELECTED_ITEMS` — conflict resolution must happen even when 0 action items selected.*
 
 **Branch-safety pre-check** — must run BEFORE `gh pr checkout` so a wrong-branch commit is impossible (per `git-commit.md` Gate 2). Verify the PR's `headRefName` is not the repo's default branch — `gh pr checkout` of a same-repo PR whose HEAD = default branch would land us on default and any later commit (Step 8) would violate Gate 2:
 
@@ -423,6 +428,8 @@ Read and execute `$_OSS_RESOLVE/modes/conflict-resolution.md`.
 
 ## Step 8: Implement action items
 
+*Skip when `SELECTED_ITEMS` is empty — jump to Step 9.*
+
 **Soft cap: 8 Codex dispatches per session.** If `SELECTED_ITEMS` has > 8 items, invoke `AskUserQuestion`: "N items selected — Codex cap is 8 per session. Split into batches?" Options: (a) Apply first 8 now, re-run for remainder · (b) Apply all [req] items only (if ≤8) · (c) Proceed anyway (sequential, may be slow).
 
 <!-- Step 8 defined in action-item-dispatch.md — see that file for phase/sub-step detail -->

plugins/oss/skills/resolve/modes/action-item-dispatch.md

@@ -23,7 +23,18 @@ trap 'rm -f "$SENTINEL"; rm -rf "$IMPL_DIR"' EXIT INT TERM
 CHALLENGE_LOG=()  # per-item records: id|evidence|suggestion|resolution
 ```
 
-`CODEX_AVAILABLE=false`: apply degradation rules from Step 1 (simple items → foundry:sw-engineer; complex items → skip with notice). Never blanket-skip all items.
+`CODEX_AVAILABLE=false`: use `change` field to route to internal agent; never blanket-skip all items.
+
+| `change` value | `IMPL_AGENT` |
+| --- | --- |
+| `code` · `refactor` · `config` · `ci` | `foundry:sw-engineer` |
+| `test` | `foundry:qa-specialist` |
+| `docs` | `foundry:doc-scribe` |
+| `style` | `foundry:linting-expert` |
+
+Complex items (effort `xhigh`, multi-file) with `CODEX_AVAILABLE=false` → skip with `⚠ codex not found — skipping item #<id> (xhigh effort). Install: /plugin marketplace add openai/codex-plugin-cc`.
+
+`--agent <name>` overrides this routing table unconditionally.
 
 > **Conflict gate**: verify all Step 5a conflict tasks `completed` before any action item. Still `pending`/`in_progress` → stop, surface list, wait. Items on unresolved conflicts compound diff.