Skip to content
Agentmart
Submit a tool

assess-pr

The /assess end-of-run offers - open a PR with the report, track the Top 3 Actions in the user's issue tracker, freeze the assessment into a CI gate, and file…

Skill by bjcoombs

IndexedApache-2.0★ 32
Write a review

Install

Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.

npx agentmarket install skill/bjcoombs-assess-pr

Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.

Get the weekly diff

The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.

We use your email only for this newsletter. Privacy

About

The /assess end-of-run offers - open a PR with the report, track the Top 3 Actions in the user's issue tracker, freeze the assessment into a CI gate, and file tool feedback. TRIGGER when the /assess orchestrator reaches the end-of-run offers; not a standalone user command.

Use cases: documents, devops

No reviews yet. Installed it? Reviews come from verified installers only. How reviews work

Permissions

What this skill can do and what it asks for, from its SKILL.md and code. Checked Oct 25, 2026.

Tools
No tools found
Secrets
None asked for
Install scripts
None found

Network hosts it talks to 1

  • codeload.github.com · found in install command

Files it touches

  • ~/.claude/skills/assess-pr

What's inside

Skill name
assess-pr
Description
The /assess end-of-run offers - open a PR with the report, track the Top 3 Actions in the user's issue tracker, freeze the assessment into a CI gate, and file tool feedback. TRIGGER when the /assess orchestrator reaches the end-of-run offers; not a standalone user command.
License
Apache-2.0
Instructions
401 lines in SKILL.md · view on GitHub

Agent Skills spec issues

  • files: referenced but missing: references/uninstall.md

Files (1)

  • SKILL.md

SKILL.md (first lines)

# Assess PR and Issues

The end-of-run half of `/assess`. The report (`.assess/assess-report.md`), the complexity heatmap, and the doc graph are written. This step runs the optional end-of-run offers - in order - then closes the loop:

1. **Open a PR** with the report (Step 5).
2. **Track the Top 3 Actions** in the user's issue tracker (Step 6).
3. **Freeze the assessment into a CI gate** (Step 6.5).
4. **Tool feedback** (Step 7) - surface anomalies and offer to file feedback.
5. **Uninstall `/assess`** (Step 8) - the escape hatch: remove everything this run wrote.

Each offer is independent (uninstall excepted - it's mutually exclusive with the write-back offers): a user can accept any subset. The step references back into the written `assess-report.md` artifact (notably the Top 3 Actions table's `Issue` column), not into the scoring prose, so it stays decoupled from how the report was produced.

## Phase 2: ask once, batched (and the non-interactive contract)

These are the **write-back phase** of the consent lifecycle (Phase 2; Phase 1 was the tool installs, Phase 3 the mutation pass - see the assess SKILL.md). Do not serialise them into back-to-back modals. Present them as **one batched, multi-select AskUserQuestion**: "Now that the report is written, which of these should I do?" with the options (open a PR, track the Top 3 Actions, freeze a CI gate, file feedback, and - the mutually-exclusive escape hatch - uninstall `/assess` from this repo), pre-filtered by feasibility:

- Drop the **PR** option when the push-capability / remote check below (Step 5) shows no direct or fork PR is possible; on a read-only target, offer the fork variant instead. When the target states it takes no contributions (Step 5's no-contributions scan), offer the **no-contributions** variant instead on a read-only target (READ / TRIAGE) or in a clone of the user's own fork: a PR inside the user's fork, never one against upstream. A user who can push to the non-fork repo itself keeps the direct PR option, with the statement quoted.
- Drop the **CI gate** option when the workflow could never run (no GitHub remote).
- Keep **issue tracking** and **feedback** always (feedback needs no repo write).
- **Uninstall** (Step 8) always appears: it removes what this run wrote. It doesn't compose with the write-back offers (no point opening a PR *and* deleting the report), so treat selecting it as "skip the others and clean up".

Then execute the accepted offers **in order** (PR → issues → gate → feedback), because the later ones cross-link to the PR the first may have opened. Uninstall, if chosen, runs instead of the others. The per-step detail below is the execution recipe for each accepted offer; the *asking* happens once, here.

**Non-interactive contract (headless / CI).** When `run-context.json .interactive` is `false`, **make no AskUserQuestion call at all** - every write-back offer is already recorded as `{type, status: "skipped", reason: "non-interactive"}` in `run-context.json .offers` (`pr`, `issue_tracking`, `ci_gate`, `feedback`, `uninstall`). Write nothing back, open no PR, create no issues, emit no gate, file no feedback, uninstall nothing; the run ends with the local `.assess/` artifacts only. A headless `/assess` completes with zero prompts.

```bash
jq '{interactive, offers}' "$REPO_ROOT/.assess/run-context.json"
```

---

## Step 5: Ask Whether to Open a PR

After writing the files, first **check whether a direct PR is even possible** before offering one. A user on `READ` or `TRIAGE` access can't push a branch to the target repo, so an unconditional "open a PR?" offer is infeasible and wastes a turn.

```bash
# Detect push capability. `gh` returns viewer fields for the current user.
# Push capability is derived from `viewerPermission` (one of ADMIN |
# MAINTAIN | WRITE | TRIAGE | READ | NONE) - GitHub's GraphQL Repository
# type has no `viewerCanPush` field, so don't request it (the CLI errors
# and the whole call returns empty, silently degrading every write-
# accessible repo to the "leave local" branch).
PUSH_INFO=$(gh repo view --json viewerPermission,viewerCanAdminister,nameWithOwner,isFork,owner 2>/dev/null || true)
PERM=$(echo "$PUSH_INFO" | jq -r '.viewerPermission // empty')
case "$PERM" in
  ADMIN|MAINTAIN|WRITE) CAN_PUSH=1 ;;
  *)                    CAN_PUSH=0 ;;
esac
# A clone of the user's own fork reports ADMIN, so CAN_PUSH alone cannot tell
# the upstream's maintainer from an outsider working in their fork.
IS_FORK=$(echo "$PUSH_INFO" | jq -r '.isFork // false')
# isFork says the repo has a parent, not whose fork it is: a collaborator with
# push access to someone else's fork also sees isFork=true and WRITE. The
# current repo counts as the user's own fork only when the viewer owns it.
VIEWER_LC=$(gh api user 2>/dev/null | jq -r '.login // empty' | tr '[:upper:]' '[:lower:]')
OWNER_LC=$(echo "$PUSH_INFO" | jq -r '.owner.login // empty' | tr '[:upper:]' '[:lower:]')
if [ "$IS_FORK" = true ] && [ -n "$VIEWER_LC" ] && [ "$VIEWER_LC" = "$OWNER_LC" ]; then
  IS_OWN_FORK=true
else
  IS_OWN_FORK=false
fi
# If the command failed (no remote, no gh, not a GitHub repo, unauthenticated),
# $PUSH_INFO is empty and $PERM stays empty - fall back to the local-branch
# flow with the reason. Never silently assume push works.

# Detect remote-to-gh redirect: GitHub follows transfers/renames, so gh may
# resolve a different owner/repo than the configured remote implies.
REMOTE_URL=$(git remote get-url origin 2>/dev/null || true)
if [ -n "$REMOTE_URL" ]; then
  # Normalise SSH and HTTPS forms (and GitHub Enterprise hosts) to owner/repo.
  # SSH:   git@host:owner/repo.git   -> owner/repo
  # HTTPS: https://host/owner/repo/  -> owner/repo
  # Strip protocol, any user@, the host (not just github.com), a trailing .git,
  # and a trailing slash - so a GHE host or a trailing slash can't masquerade as
  # a redirect.
  REMOTE_SLUG=$(echo "$REMOTE_URL" \
    | sed -e 's|^[a-zA-Z]*://||' -e 's|^[^@/]*@||' -e 's|^[^/:]*[:/]||' -e 's|\.git$||' -e 's|/$||')
else
  REMOTE_SLUG=""
fi
GH_SLUG=$(echo "$PUSH_INFO" | jq -r '.nameWithOwner // empty')
# Compare case-insensitively: GitHub treats owner/repo as case-insensitive, so a
# pure case difference is not a redirect and must not emit a spurious notice.
REMOTE_SLUG_LC=$(echo "$REMOTE_SLUG" | tr '[:upper:]' '[:lower:]')
GH_SLUG_LC=$(echo "$GH_SLUG" | tr '[:upper:]' '[:lower:]')
if [ -n "$REMOTE_SLUG" ] && [ -n "$GH_SLUG" ] && [ "$REMOTE_SLUG_LC" != "$GH_SLUG_LC" ]; then
  REDIRECT_NOTICE="Note: origin (\`$REMOTE_SLUG\`) redirects to \`$GH_SLUG\`; the offers below target the redirected repo."
else
  REDIRECT_NOTICE=""
fi
```

Then check whether the target turns away outside contributions. A fork PR against a repo whose README or CONTRIBUTING says "do not send a pull request" is noise for its maintainers, so the scan runs before any upstream offer is framed:

```bash
# no-contributions scan: start
# Sets NO_CONTRIBUTIONS (1/0) plus the first matching file (relative to the
# repo root) and line, so the offer can quote the statement back to the user.
NO_CONTRIBUTIONS=0
NO_CONTRIBUTIONS_SOURCE=""
NO_CONTRIBUTIONS_STATEMENT=""
# The pattern is built from named parts so every branch shares one noun and one
# boundary. _nc_noun ends at a non-letter (_nc_end), so "prs?" cannot match the
# start of "promise", "process" or "private"; the passive branch also needs a
# non-letter before the noun. _nc_neg has no leading boundary on purpose: it
# must match the tail of "cannot".
_nc_neg="(not|n't|n’t|unable to)"
_nc_noun="(contributions|pull requests?|prs?)"
_nc_end="([^a-z]|$)"
# Three refusal phrasings:
#   accept:     "does not accept (community) contributions", "not accepting PRs"
#   imperative: "do not send a pull request" - the clause must end right after
#               the noun, so "do not open a PR for typo fixes" (a house rule
#               for contributors, not a refusal) stays out
#   passive:    "pull requests are not accepted"
_nc_accept="${_nc_neg} ([a-z-]+ )?accept(ing)? ([a-z-]+ )?${_nc_noun}${_nc_end}"
_nc_imperative="${_nc_neg} (send|open|submit) (a |any |us )?${_nc_noun}[[:space:]]*([.;!)]|$)"
_nc_passive="(^|[^a-z])${_nc_noun} (are|is|will) not (be )?accepted"
_nc_re="${_nc_accept}|${_nc_imperative}|${_nc_passive}"

From bjcoombs/ai-native-toolkit, Apache-2.0.

More from bjcoombs/ai-native-toolkit

Signals

Stars
32
Forks
5
Open issues
7
Last commit (this folder)
Sep 19, 2026 (1 month ago)
Last release
Oct 11, 2026 (18 days ago)

From GitHub (fetched Oct 24, 2026).

Details

License
Apache-2.0
First seen
Oct 24, 2026
Last checked
Oct 24, 2026

More from bjcoombs

Cite this page

assess-pr skill. Agent Mart. https://agentmart.dev/skills/bjcoombs-assess-pr (last updated 2026-10-24)