fix(skills): pr-review-fixer gh-missing fallback by chiptus · Pull Request #503 · chiptus/UpLine

chiptus · 2026-09-16T06:53:03Z

Split off from #498, which had this bundled unrelated to the triage-labels/git-conventions docs work there.

pr-review-fixer hard-depended on the gh CLI, which isn't installed in remote/cloud sessions (GitHub access there goes through mcp__github__* tools instead) — confirmed by a real /pr-review-fixer failure. Adds a gh-availability check and an MCP-based fallback path for fetch/resolve/reply, disclosed to its own gh-missing.md (only loaded when gh turns out to be missing) rather than inlined for every reader.

Verification

In a session without the gh CLI, /pr-review-fixer now checks command -v gh first and fetches/resolves/replies via mcp__github__* tools instead of failing on command not found — reproduced live against this repo's own PR docs(agents): triage groups, git conventions #498.
In a session with gh, the original script-based path is unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

Generated by Claude Code

Documents the five canonical triage-role labels as a mutually-exclusive Linear label group, separates lifecycle markers (agent:wip/agent:pr/epic) from that group, and consolidates branch-naming + commit/PR title format into docs/git-conventions.md, linked from CLAUDE.md and autonomic-issues.md. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

Reorders git-conventions.md so the shared <type> is defined once, before its first use; front-loads the CLAUDE.md pointer's trigger word; and adds the missing label-group setup step to the autonomic pipeline's setup checklist (single source of truth, referenced from triage-labels.md). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

New user-invoked skill that runs setup-matt-pocock-skills to pick this repo's issue tracker (and its triage labels/domain docs), then reuses that same tracker choice to scaffold docs/agents/autonomic-issues.md from a GitHub- or Linear-shaped template — never asking the tracker question twice. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

Replaces the two lifecycle labels with a single `agent` label whose stage is read off the issue's native Linear status (In Progress while claimed, In Review once the PR is open) instead of a second label. Retired agent:wip/agent:pr in Linear (team UPL) and migrated the one live issue carrying agent:pr (UPL-7, open PR #463) to the new scheme. Also updates the setup-chiptus-env Linear template to match; the GitHub template is unchanged since GitHub issues have no native per-issue status to borrow. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

…arity Removes a leftover "PR-side label" mention in autonomic-issues.md now that GitHub PRs carry no mirrored label. Adds a note to the setup-chiptus-env Linear template (not the repo docs, which are tracker-specific) explaining that GitHub's native PR state (draft/ready/merged/closed) already parities Linear's issue status, so future setups shouldn't invent a matching status label on the PR side. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

The PR-cap paragraph claimed counting stayed correct "even if the label is forgotten" — true under the old agent:pr-only count, false now that the count requires both --label agent and --status "In Review" together (a forgotten label does drop the issue from the count; this is intentional, since it's what lets a maintainer's own manually-opened PR skip the cap). Also corrects issue-tracker.md's `issues update --state` to the actual verified CLI flag, `--status`, which is what the new claim/PR-review steps already used correctly. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

setup-chiptus-env now checks for setup-matt-pocock-skills before invoking it, installing the mattpocock/skills set via `npx skills` when missing, instead of assuming it's already there. Also adds a step to redirect domain docs (CONTEXT.md, ADRs) to an external folder for repos that can't keep them in-repo (e.g. Portainer, where architecture/ customer docs must stay out of the public repo) — rewriting docs/agents/domain.md's pointer rather than touching setup-matt-pocock-skills' own templates. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

… docs Generalizes the external-docs step from just CONTEXT.md/ADRs to the entire docs/agents/ folder (issue tracker, triage labels, autonomic pipeline, domain consumer rules), matching the Portainer pattern: the external root mirrors the in-repo layout exactly, so the docs' own mutual cross-references need no rewriting — only CLAUDE.md/AGENTS.md's entry pointers and the Routine prompts need to resolve a .git/-local pointer file to the external root. Also flags the real gap this creates for cloud Routine firings, which work off a fresh clone with no access to this machine's .git/ contents. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

…only An env var (e.g. AGENTS_DOCS_REPO) holding a separate docs repo's git URL is now the default pointer mechanism, since it's settable in a Routine's own environment_variables and so actually solves the cloud- portability gap the .git/-local file couldn't. The .git/ file stays as a lighter option for a solo local setup with no Routines involved, but the skill now asks the user to choose explicitly between the two rather than defaulting silently, since the choice decides whether the autonomic pipeline can read the docs at all. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

…nsumers CLAUDE.md's pointers stay plain ("See docs/agents/X.md", no conditional phrasing) since it's always read locally regardless of where docs live and per-pointer resolution text would duplicate the same logic at every call site. Instead, step 3 now has the skill grep for whichever mattpocock skills actually read CONTEXT.md/docs/adr/docs/agents directly (domain-modeling, triage, code-review, etc.) and prepend the external-docs check to each of those, plus the two Routine prompts — the actual code paths that need to know, once, in one place per file. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

autonomic-issues-github.md and autonomic-issues-linear.md were ~90% identical (triage bar, quality gates, guardrails, notifications, both routine prompts verbatim) with only the label/status mechanism, CLI commands, and a few paragraphs differing per tracker — real duplication of a single source of truth. Merged into one autonomic-issues.md with the differing paragraphs wrapped in  /  marker pairs, positioned right next to their counterpart; filling the template is now "delete the blocks for the tracker you didn't pick." Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

… fragments The interleaved-marker design didn't scale: every section grew another block per tracker, so a third tracker would mean editing every section of an already-written file. Replaces it with a skeleton (autonomic-issues.md, 100% tracker-invariant prose) plus small per-tracker fragment files (tracker-github.md, tracker-linear.md) that each supply the same five named slots (INTRO, SHARED_STATE, INTAKE_QUERY, FIX_FIRING_STEPS, SETUP_ITEMS). Adding a tracker is now "write one new fragment file against these five slots" — the skeleton never changes. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

The Setup checklist's items 3-4 lived in the skeleton after the fragment-supplied items 1-2, relying on markdown list continuation across a template substitution boundary — fragile in general, and prettier's markdown reflow already collapsed it into one broken paragraph on the prior commit. SETUP_ITEMS now owns the whole 1-4 list in each fragment (2 short duplicated lines) instead. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

…l table Both prior designs (interleaved markers, then skeleton+fragments) kept duplicating full paragraphs per tracker when only a handful of facts actually differ: how "claimed"/"in review" are represented, how priority works, how a PR declares its issue link. Rewrites the whole doc in tracker-agnostic language (triage rubric, quality gates, guardrails, bail/fail handling, notifications, routine prompts never mention a tracker) referring back to one small "Tracker specifics" table for the load-bearing facts only — not a CLI reference, which the agent already gets from the tracker's own usage/--help. Adding a tracker is now one new table column, not a new file. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

GitHub and Linear are columns in the table, not rows; a new tracker adds a column. Matches the wording already used in SKILL.md. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

Missed applying the earlier agent-label-plus-native-state decision to GitHub's own column: it still described the old agent:wip -> agent:pr swap. A GitHub PR's own draft/ready-for-review/merged state already tells you claimed vs in-review once a PR exists, same reasoning as Linear's native status — so GitHub gets one `agent` label too, applied at claim and never swapped. Updated the two callers that assumed a literal marker-move action (Fix firing step 6, the manual-PR coexistence note) to describe the inferred state instead. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

The "don't add a matching status label on the GitHub PR side" note was sitting in the Linear cell even though its subject is GitHub's own PR state; moved to the GitHub cell (referencing the Linear cell for contrast) instead of being explained from the other tracker's row. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

…s missing The skill's fetch/resolve/reply steps hard-depended on the gh CLI, which isn't installed in this remote session (GitHub access here goes through mcp__github__* tools instead) — confirmed by the failed fetch-review-threads.sh run. Adds an explicit gh-availability check at the top of Phase 1 and an MCP-based path alongside each gh-dependent step (fetch via pull_request_read, resolve via resolve_review_thread, reply via add_reply_to_pull_request_comment / add_issue_comment) that reconstructs the same data shape, so the rest of the skill is unaffected by which path ran. Note: this file is npx skills-managed (mattpocock/skills) — a future bare reinstall would overwrite this fix back out. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

Linear's mutual-exclusivity feature kept the five canonical triage-role labels enforced structurally, but nothing said what covers that on GitHub, which has no equivalent group feature. Adds a table row and a setup-checklist clause: GitHub relies on the triage skill's own discipline (remove any other triage-role label before applying a new one) instead of the tracker enforcing it. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

…r's grouped look Confirmed via linearis that Linear's own label names stay plain (ready-for-agent, not triage/ready-for-agent) — the grouped display comes from each label's parent pointing at the triage group label, which Linear's UI renders hierarchically. GitHub has nothing equivalent, so the closest substitute is naming the five labels with a literal triage/ prefix. Documented in the Tracker specifics table (with the Linear cell clarifying the contrast) and suggested as a default during setup-matt-pocock-skills' own label-naming question. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

- Disclosed step 3's external-docs mechanics (pointer options, layout, the grep-and-patch recipe for consumer skills) into a new external-docs.md, reached only on the "external" branch most repos never take, instead of sitting inline in the main procedural flow every reader sees. - Trimmed a duplicated explanation in the Tracker specifics table: the GitHub "Lifecycle labels" row restated the draft/ready/merged mapping the "Claimed / in review" row above it already gives: now points back at it instead. - Fixed a stale "the templates generate" (plural, left over from the since-collapsed multi-file design). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

This session iterated the autonomic-issues.md template through four different shapes reactively; worth a slower discussion session before more incremental edits. Captures the open questions (tracker-table generality, external-docs relocation being unvalidated, npx skills unverified, the triage/ prefix suggestion being unvalidated) rather than restating the diffs, which live in this branch's commit history. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

Split out to #499, since it still needs a discussion pass before it's done (see that PR's handoff doc) and shouldn't hold up these settled repo-docs changes. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

The previous fix wrapped the fetch script in "gh available: run ...", but that used the !`command` auto-exec markdown syntax, which the harness pre-executes unconditionally at skill-load time -- before the model ever reads the conditional prose around it. That's exactly what the reported /pr-review-fixer failure showed: the script still ran (and failed on gh: command not found) despite the branch text saying to check first. Removes the auto-exec syntax so the agent runs the check and the script itself via the Bash tool, where the branch can actually take effect. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

…ts own file Per review feedback on PR #498: the "gh missing" branch (Phase 1's MCP fetch recipe, Phase 4's resolve/reply mechanics) was inline in every reader's path even though only some sessions need it. Split into gh-missing.md, pointed to from Phase 1 and Phase 4 only where the gh CLI turns out to be unavailable -- same progressive-disclosure move already applied to setup-chiptus-env. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

This branch keeps only the pr-review-fixer gh-fallback fix, unrelated to the triage-labels/git-conventions docs work on docs/triage-labels-git-conventions / PR #498. The settled repo docs stay there instead. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

vercel · 2026-09-16T06:53:09Z

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
upline	Ready	Preview	Sep 16, 2026 8:08am UTC

Split out to #503, unrelated to this branch's triage-labels/git-conventions docs work. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

Per Copilot review on PR #498 (posted after the file had already moved here): allowed-tools grants no git command, only "command -v gh" and the GitHub MCP calls, so "owner/repo from the git remote" was an instruction the skill has no permission to carry out. Points at the session's own repo context instead. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

Copilot

🟡 Changes recommended

Unresolved findings block reliable PR discovery and MCP review-thread handling.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Adds an MCP fallback for pr-review-fixer when the gh CLI is unavailable.

Changes:

Detects gh availability and preserves the existing CLI path.
Documents MCP-based fetching, resolving, and replying.

File summaries

File	Summary and findings
`.agents/skills/pr-review-fixer/SKILL.md`	Adds fallback routing. Moderate (1 vote): the fallback cannot discover the owner, repository, and PR number without additional lookup support or required inputs.
`.agents/skills/pr-review-fixer/gh-missing.md`	Documents the MCP workflow. Moderate (1 vote): no supported PR-number discovery path. Critical (1 vote): returned comment data and IDs do not match the review-thread shape required by later phases, including line 15.

Review details

Suppressed comments (3)

.agents/skills/pr-review-fixer/SKILL.md:10

The fallback cannot reliably identify the PR it is supposed to read. pull_request_read needs an owner/repo and PR number, but this skill's allowlist only permits command -v gh (not git remote, branch inspection, or a GitHub lookup). gh-missing.md then tells the agent to derive those values from the current branch, so a normal /pr-review-fixer invocation with no PR number cannot reach Phase 1. Allow the needed lookup tool(s), or require the caller to provide owner/repo/number before invoking the fallback.

allowed-tools: Bash(${CLAUDE_SKILL_DIR}/scripts/fetch-review-threads.sh) Bash(${CLAUDE_SKILL_DIR}/scripts/resolve-thread.sh *) Bash(gh pr comment *) Bash(command -v gh) mcp__github__pull_request_read mcp__github__resolve_review_thread mcp__github__add_reply_to_pull_request_comment mcp__github__add_issue_comment

.agents/skills/pr-review-fixer/gh-missing.md:16

Phase 4 requires comments[].html_url to derive the numeric parent comment ID, but Phase 1's fallback contract only preserves author and body for each comment. If the agent follows this normalization, inline replies have no parent ID and add_reply_to_pull_request_comment cannot be called. Explicitly retain html_url (or the numeric comment ID) in the Phase 1 shape.


- `method: get_review_comments` → review threads. Each has `id` (the GraphQL

.agents/skills/pr-review-fixer/gh-missing.md:11

Unlike the gh pr view script, this fallback has no way to determine the current PR number: a git remote only supplies owner/repo, while pull_request_read requires a pullNumber, and no PR lookup or branch-to-PR mapping is provided. A normal invocation that supplies only the current branch therefore cannot make the first MCP call. Add a supported PR-number discovery path (or require and document the PR number as input) before using this fallback.

Reconstruct the same `{threads, reviews, issueComments}` shape from
`mcp__github__pull_request_read`. This skill's `allowed-tools` grants no git
command, so don't try to derive owner/repo/PR number by shelling out — use

Files reviewed: 2/2 changed files
Comments generated: 1
Review effort level: Lite

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

+- `method: get_review_comments` → review threads. Each has `id` (the GraphQL
+  thread node ID — this is what `resolve_review_thread` below needs, keep it),
+  `is_resolved`, `path`, `line`, and `comments[]` with `author`/`body`. Keep
+  only `is_resolved == false` — the tool doesn't filter this for you the way
+  the script's `jq` does.


github-actions · 2026-09-16T06:59:59Z

Playwright test results

65 passed

Details

65 tests across 21 suites
1 minute, 19 seconds
36a6b3a

The "reads {threads, reviews, issueComments} the same way" sentence was stated once in SKILL.md's Phase 1 already, ahead of the pointer into gh-missing.md, so repeating it there was redundant. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01776GRzBiT1RkWHD7vyctrp

claude added 27 commits September 13, 2026 06:42

fix(skills): row-count mention was stale after adding a fifth row

02e2acf

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01HFNqcr5d2EN6T6TkTc8WBd

Copilot AI lite review requested due to automatic review settings September 16, 2026 06:53

Copilot started reviewing on behalf of chiptus September 16, 2026 06:53 View session

vercel Bot deployed to Preview September 16, 2026 06:53 View deployment

vercel Bot deployed to Preview September 16, 2026 06:54 View deployment

Copilot AI reviewed Sep 16, 2026

View reviewed changes

vercel Bot deployed to Preview September 16, 2026 08:08 View deployment

chiptus mentioned this pull request Sep 16, 2026

fix(pr-review-fixer): add gh-missing fallback via mcp__github__ tools chiptus/skills#5

Open

Navigation Menu

fix(skills): pr-review-fixer gh-missing fallback - #503

fix(skills): pr-review-fixer gh-missing fallback#503
chiptus wants to merge 29 commits into
mainfrom
fix/pr-review-fixer-gh-fallback

chiptus commented Sep 16, 2026

Uh oh!

vercel Bot commented Sep 16, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

github-actions Bot commented Sep 16, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

chiptus commented Sep 16, 2026

Verification

Uh oh!

vercel Bot commented Sep 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Suppressed comments (3)

Uh oh!

github-actions Bot commented Sep 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Playwright test results

Details

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

vercel Bot commented Sep 16, 2026 •

edited

Loading

github-actions Bot commented Sep 16, 2026 •

edited

Loading