Skip to content

Commit 64b695f

Browse files
jantmanclaude
andcommitted
Rebuild the Claude workflows on the privatepuppet template
Both Claude workflows are replaced with the shared versions that came out of jantman/privatepuppet, so that all ten repositories running these run the same file. `claude-pr-review.yml` is now byte-identical across every repository; `claude-mention.yml` differs only in its `if:` guard and one branch name. The problem these fix is that a broken run here is indistinguishable from a quiet one. The action denies a tool call, the agent improvises around the gap, ends its turn and exits 0 -- and with `display_report` off, the only evidence is a `permission_denials_count` buried in a streamed result object that never reaches the log. That is how `/code-review:code-review` ran for months without reviewing anything: it is a Skill, and `Skill` was not in the allowlist. What changes: - `Skill`, `Task`, `Write` and `TodoWrite` are in the review allowlist. Agent mode does not merge a base tool set into `--allowedTools`, so anything absent there is absent from the run, and `/code-review:code-review` is itself a Skill. - The enumerated `Bash(...)` allowlists are gone. The review plugin fans out subagents that reach for sed/awk/jq/diff, and every unlisted command is a silent denial rather than an error. - `display_report` and `show_full_output` are on, and each run uploads its transcript and session logs as an artifact. The action otherwise throws both away with the runner. - The agent no longer posts to GitHub. It writes its findings to a JSON file and a later step turns them into exactly one GitHub review, with the findings as batched inline comments and the run's duration, turn count and cost in the body. A run that decides to skip still posts a one-line review, because a skip that should not have happened is otherwise invisible. - Duration, turns, cost and permission denials go in the job summary, and a denial or an `is_error` result now fails the check. - `mcp__github_ci__*` and `actions: read` let the reviewer read the CI results for the commit it is reviewing instead of guessing why a check went red. - The review job skips pull requests from forks. A fork run gets no secrets, so the action would fail red on someone else's contribution. Because this repository is public, `claude-mention.yml` additionally requires the triggering author's `author_association` to be OWNER, MEMBER or COLLABORATOR. That job has `contents: write` and blanket `Bash`, so without the check any passer-by who types `@claude` gets an agent run billed to the maintainer's account. The field is read per-event on purpose: on `issue_comment`, `github.event.issue.author_association` describes the issue's opener rather than the commenter, so reading it there would let a stranger trigger a run merely by commenting on a maintainer's own issue. Checked with actionlint and by bash -n on every `run:` block. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
1 parent 6342511 commit 64b695f

4 files changed

Lines changed: 877 additions & 133 deletions

File tree

‎.github/workflows/claude-code-review.yml‎

Lines changed: 0 additions & 44 deletions
This file was deleted.
Lines changed: 196 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,196 @@
1+
name: Claude Code
2+
3+
# Reactive: does nothing until someone writes @claude. Pairs with
4+
# claude-pr-review.yml, which reviews every PR unprompted.
5+
#
6+
# >>> DO NOT MERGE THESE TWO INTO ONE JOB. <<<
7+
# They look redundant and are not. Supplying a `prompt` puts the action in AGENT
8+
# mode for every event it sees, so a single job carrying both triggers would
9+
# answer @claude comments in agent mode: no PR context, no tracking comment, and
10+
# nothing posted back. This file deliberately supplies no prompt, which is what
11+
# keeps it in TAG mode. Deleting one as a duplicate of the other has silently
12+
# broken @claude before.
13+
#
14+
# NOTE: `issue_comment` workflows always run from the default branch, so edits to
15+
# this file do nothing until they are on `main`.
16+
#
17+
# The 👀 reaction on an @claude comment comes from the GitHub App acknowledging the
18+
# mention. It is NOT evidence that anything ran -- the work happens here.
19+
20+
on:
21+
issue_comment:
22+
types: [created]
23+
pull_request_review_comment:
24+
types: [created]
25+
issues:
26+
types: [opened, assigned]
27+
pull_request_review:
28+
types: [submitted]
29+
30+
jobs:
31+
claude:
32+
# Two conditions, both required, and they must be checked per-event.
33+
#
34+
# 1. author_association of the person who ACTUALLY TRIGGERED IT. This repository is
35+
# public and this job has `contents: write` and blanket `Bash`, so without this
36+
# check any passer-by who types @claude gets an arbitrary-code-execution agent
37+
# run on the maintainer's OAuth token. The pairing matters: on `issue_comment`,
38+
# `github.event.issue.author_association` is the issue's *opener*, not the
39+
# commenter, so checking that field here would let a stranger trigger a run
40+
# merely by commenting on a maintainer's own issue. OWNER covers the repository
41+
# owner; COLLABORATOR/MEMBER cover anyone later given push access.
42+
# 2. the @claude mention itself.
43+
if: |
44+
(github.event_name == 'issue_comment' &&
45+
contains(github.event.comment.body, '@claude') &&
46+
contains(fromJSON('["OWNER","MEMBER","COLLABORATOR"]'), github.event.comment.author_association)) ||
47+
(github.event_name == 'pull_request_review_comment' &&
48+
contains(github.event.comment.body, '@claude') &&
49+
contains(fromJSON('["OWNER","MEMBER","COLLABORATOR"]'), github.event.comment.author_association)) ||
50+
(github.event_name == 'pull_request_review' &&
51+
contains(github.event.review.body, '@claude') &&
52+
contains(fromJSON('["OWNER","MEMBER","COLLABORATOR"]'), github.event.review.author_association)) ||
53+
(github.event_name == 'issues' &&
54+
(contains(github.event.issue.body, '@claude') || contains(github.event.issue.title, '@claude')) &&
55+
contains(fromJSON('["OWNER","MEMBER","COLLABORATOR"]'), github.event.issue.author_association))
56+
runs-on: ubuntu-latest
57+
permissions:
58+
contents: write
59+
pull-requests: write
60+
issues: write
61+
id-token: write
62+
actions: read # Required for Claude to read CI results on PRs
63+
steps:
64+
- name: Checkout repository
65+
uses: actions/checkout@v4
66+
with:
67+
fetch-depth: 1
68+
69+
- name: Run Claude Code
70+
id: claude
71+
uses: anthropics/claude-code-action@v1
72+
with:
73+
claude_code_oauth_token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
74+
75+
# This is an optional setting that allows Claude to read CI results on PRs.
76+
# It is also what makes the three `mcp__github_ci__*` tools available; tag
77+
# mode adds them to the allowlist on its own, so they are not listed below.
78+
additional_permissions: |
79+
actions: read
80+
81+
# No `prompt:` on purpose -- see the header. That is what keeps this in tag
82+
# mode, where Claude performs the instructions in the comment that tagged it.
83+
84+
# Without these the log shows only "init" and "result". On a public
85+
# repository the job log and the artifacts below are world-readable, so this
86+
# is a deliberate trade: without them, a run that was starved by a denied
87+
# tool call is indistinguishable from one that had nothing to do.
88+
display_report: true
89+
show_full_output: true
90+
91+
# >>> Do not go back to an enumerated command allowlist. <<<
92+
# Earlier versions of this file listed individual Bash commands (puppet-lint,
93+
# rake spec, ./bin/run_tests.sh, nox, poetry, ...). Anything unlisted is a
94+
# SILENT denial: the run goes green having done nothing, which is exactly how
95+
# the PR-review job failed before it was rebuilt. Skill and Task matter for
96+
# the same reason -- a skill or a subagent that cannot be invoked is not an
97+
# error, just an absence.
98+
#
99+
# Unlike the review job, this one can push, so it is the more powerful of the
100+
# two. What keeps blanket Bash acceptable is who can reach it at all -- see
101+
# the `if:` above and the comment on it.
102+
#
103+
# In tag mode this list is MERGED with the action's own base tools (Read/Glob/
104+
# Grep/LS, the comment-update tool, the CI tools, and git add/commit/push/rm);
105+
# it does not replace them. Naming an `mcp__github_*` tool here is also what
106+
# installs the corresponding MCP server, which is why the inline-comment tool
107+
# is listed even though this job rarely reaches for it.
108+
#
109+
# WebSearch and WebFetch are DISALLOWED by default in tag mode; naming them
110+
# here is what lifts that. They are worth having -- "@claude why does this
111+
# library do X" is most of what the mention job gets asked -- and the `if:`
112+
# above is what makes it an acceptable trade, since an agent that can fetch
113+
# arbitrary URLs and also push is only reachable by someone with push access.
114+
claude_args: |
115+
--allowedTools "mcp__github_inline_comment__create_inline_comment,Read,Glob,Grep,Edit,Write,Task,Skill,TodoWrite,WebSearch,WebFetch,Bash"
116+
117+
# The action discards its transcript with the runner. Without this there is no way
118+
# to explain a green run that did nothing.
119+
- name: Collect Claude logs
120+
if: always()
121+
env:
122+
EXECUTION_FILE: ${{ steps.claude.outputs.execution_file }}
123+
run: |
124+
mkdir -p "$RUNNER_TEMP/claude-logs"
125+
if [ -n "$EXECUTION_FILE" ] && [ -f "$EXECUTION_FILE" ]; then
126+
cp "$EXECUTION_FILE" "$RUNNER_TEMP/claude-logs/execution-output.json"
127+
fi
128+
if [ -d "$HOME/.claude/projects" ]; then
129+
cp -r "$HOME/.claude/projects" "$RUNNER_TEMP/claude-logs/sessions"
130+
fi
131+
find "$RUNNER_TEMP/claude-logs" -type f -printf '%s\t%p\n' | sort -rn
132+
133+
- name: Upload Claude logs
134+
if: always()
135+
uses: actions/upload-artifact@v7
136+
with:
137+
name: claude-mention-logs-${{ github.run_id }}-${{ github.run_attempt }}
138+
path: ${{ runner.temp }}/claude-logs/
139+
if-no-files-found: warn
140+
retention-days: 14
141+
142+
# Same guard as claude-pr-review.yml: a denied tool call is how one of these
143+
# jobs passes while having done nothing. There is no comment to annotate here
144+
# -- in tag mode the action maintains its own tracking comment -- so the
145+
# numbers go to the job summary and a denial fails the check.
146+
- name: Summarise the run and fail on permission denials
147+
if: always()
148+
env:
149+
EXECUTION_FILE: ${{ steps.claude.outputs.execution_file }}
150+
run: |
151+
set -uo pipefail
152+
153+
if [ -z "${EXECUTION_FILE:-}" ] || [ ! -s "$EXECUTION_FILE" ]; then
154+
echo "::warning title=Claude did not run::No execution file was produced, so the agent never started. Expected if this PR edits .github/workflows/claude-*.yml (workflow validation skip)."
155+
exit 0
156+
fi
157+
158+
RESULT=$(jq -c '[.[] | select(.type == "result")] | last // empty' "$EXECUTION_FILE")
159+
if [ -z "$RESULT" ]; then
160+
echo "::warning title=No result block::The execution file exists but contains no result object."
161+
exit 0
162+
fi
163+
164+
DURATION_MS=$(jq -r '.duration_ms // 0' <<<"$RESULT")
165+
TURNS=$(jq -r '.num_turns // 0' <<<"$RESULT")
166+
COST=$(jq -r '.total_cost_usd // 0' <<<"$RESULT")
167+
DENIALS=$(jq -r '.permission_denials_count // 0' <<<"$RESULT")
168+
IS_ERROR=$(jq -r '.is_error // false' <<<"$RESULT")
169+
170+
HUMAN=$(awk -v ms="$DURATION_MS" 'BEGIN {
171+
s = int(ms / 1000)
172+
if (s < 60) { printf "%ds", s } else { printf "%dm %ds", int(s / 60), s % 60 }
173+
}')
174+
COST_FMT=$(awk -v c="$COST" 'BEGIN { printf "%.4f", c }')
175+
176+
{
177+
echo "### Claude (@mention)"
178+
echo
179+
echo "| | |"
180+
echo "|---|---|"
181+
echo "| Duration | $HUMAN |"
182+
echo "| Turns | $TURNS |"
183+
echo "| Cost | \$$COST_FMT |"
184+
echo "| Permission denials | $DENIALS |"
185+
} >> "$GITHUB_STEP_SUMMARY"
186+
187+
STATUS=0
188+
if [ "$DENIALS" != "0" ]; then
189+
echo "::error title=Claude was denied $DENIALS tool call(s)::A denied tool means the run was starved, even though it exited cleanly. Add the missing tool to --allowedTools. See the claude-mention-logs-* artifact."
190+
STATUS=1
191+
fi
192+
if [ "$IS_ERROR" = "true" ]; then
193+
echo "::error title=Claude reported an error::The run's result block has is_error=true."
194+
STATUS=1
195+
fi
196+
exit $STATUS

0 commit comments

Comments
 (0)