[eatmoreduck]

为什么改

同是控制台 Auto-review 的 dispatch「Review #N」模式，panta-log（官方模板原样）的评审署名是 pullfrog[bot]，本仓库却是 github-actions[bot]。逐项对照后差异有两处，都让 GITHUB_TOKEN 抢走了评审提交权：

1. Run agent 步骤多塞了 env.GH_TOKEN: ${{ github.token }}
2. job permissions 给了 issues: write + pull-requests: write——GITHUB_TOKEN 有评审写权限，action 就自己提交

panta-log 官方模板两处都没有：GITHUB_TOKEN 只能读代码，action 无路可走，只能由服务端（OIDC 认证）以 pullfrog[bot] App 身份提交评审。

改了什么

• 【删除】Run agent 步骤的 env.GH_TOKEN（5ef8609）
• 【修改】job permissions 收敛为 id-token: write + contents: read，与官方模板一致（e4ee4f3）
• 【修改】文件头凭据模型 + 两处注释，记录实测对照结论
• 【保留】Checkout 步骤的 GH_TOKEN（只读拉代码，与评审提交无关）、prompt 契约、progress_comments: disabled

怎么测的 + 一个机制说明

workflow_dispatch 的 workflow 定义取自 dispatch 的 ref（master）——第一轮在 PR #137 自身上的验证不算数（07:37 的评审跑的还是 master 上的旧文件，见 run 36227320805）。所以正确验证流程是：

1. 合并本 PR（改动进入 master）
2. 开一个后续测试 PR → Auto-review 照常派发 → 看该评审署名是否为 pullfrog[bot]

yaml.safe_load 已验证：job permissions 仅 contents + id-token，Run agent env 无 GH_TOKEN。

🤖 Generated with Claude Code

[github-actions]

Caution

Merging this as-is breaks .github/workflows/pullfrog.yml. Removing the step's only env key leaves env: as an empty mapping (YAML null), which GitHub rejects with Invalid workflow file … Unexpected value '' — every auto-review dispatch would fail to parse before the agent even starts.

Reviewed changes

• Dropped env.GH_TOKEN from the Run agent step — relies on the action's OIDC / App-token path so reviews are submitted as pullfrog[bot]; the token-free direction matches Pullfrog's documented model (id-token: write + ASKPASS git auth), with Checkout keeping its own GH_TOKEN.
• Rewrote the header credential note and the step comment — records that github.token now belongs only to Checkout and warns against re-adding GH_TOKEN.

ℹ️ Header still says reviews post as github-actions[bot]

The file header at lines 5-6 still reads「评审结论以 github-actions[bot] 名义发 review」, which the change itself now contradicts (the new lines 23-24 say the opposite). It is outside every diff hunk, so it is not inline-annotated here — worth fixing in this same doc pass, since the whole point of the header edit is to stop the identity from drifting back.

Technical details

# Stale reviewer-identity sentence in `pullfrog.yml` header

## Affected sites
- `.github/workflows/pullfrog.yml:5-6` —「并自动填好 prompt，评审结论以 github-actions[bot] 名义发 review」describes the pre-change behavior.

## Required outcome
- The header states the reviewer identity this workflow now produces (`pullfrog[bot]`), or drops the identity claim entirely.

## Suggested approach
- Update the sentence to `pullfrog[bot]` and, optionally, cross-reference the step comment at line 85.

ℹ️ This PR cannot validate itself via its own auto-review

Pullfrog auto-review dispatches run against the default branch, so the review landing on this PR executes master's pullfrog.yml — the one that still sets GH_TOKEN. The run for this PR (Review #137 [jauax]) is pinned to head_sha 6db7fee on master, confirming the pre-change file is what runs. Expect the auto-review on this PR to stay github-actions[bot] even after this fix; the real check has to happen on a run dispatched after merge.

Technical details

# Verification plan runs the wrong workflow revision

## Affected sites
- PR body "怎么测的" —「本 PR 即验证载体…直接看本 PR 评审的署名」.
- `.github/workflows/pullfrog.yml` is only reachable via `workflow_dispatch`, so a PR merged ref never executes here.

## Evidence
- Workflow-run list: every auto-review run has `head_branch: master`; `Review #137` is at `head_sha 6db7fee` (master, pre-merge).
- `docs.pullfrog.com/pr-reviews.md`: "Pullfrog runs … are dispatched against your default branch — so the workflow run never attaches to the pull request."

## Required outcome
- Verify on a post-merge run: merge, then trigger a review on another PR (or a fresh `workflow_dispatch`) and confirm the review author is `pullfrog[bot]` (compare against #135, authored by `github-actions[bot]`).

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using DeepSeek Flash ｜ 𝕏}

[github-actions]

Removing the only key leaves this step's env: with no entries, so the value is YAML null — and GitHub rejects an empty env: mapping with Invalid workflow file … (Line: 84, Col: 9): Unexpected value '', which invalidates the whole workflow. yaml.safe_load returning None confirms the null but not GitHub's grammar; delete the env: key entirely (the comments can move above the step).

Technical details

# Empty `env:` mapping invalidates the workflow

## Affected sites
- `.github/workflows/pullfrog.yml:84` — `env:` has only comment children (lines 85-90), so its value is `null`.

## Evidence
- `actionlint .github/workflows/pullfrog.yml` → `84:13: expecting a single ${{...}} expression or mapping value for "env" section, but found plain text node`.
- GitHub: an empty `env:` block is an invalid workflow file — `Unexpected value ''` (see openvinotoolkit/openvino#37043, which cites the same failure for an empty `env:` at the workflow level; the step-level grammar is identical).
- `yaml.safe_load(open('.github/workflows/pullfrog.yml'))` returns `None` for this step's `env`.

## Required outcome
- The `Run agent` step declares no `env` key at all; `Checkout` keeps its own `GH_TOKEN`.

## Suggested approach
- Delete line 84 and lift the three comments above the `uses:` / `with:` block; no behavior is lost because the step intentionally has no env vars.

[github-actions]

Addressed in 2fec60f — the env: key is deleted (comments lifted to step level); actionlint now passes clean.

  ^{｜ View workflow run ｜ via Pullfrog ｜ Using DeepSeek Flash ｜ 𝕏}

[github-actions]

Caution

The two new commits are correct in direction, but this PR still cannot merge — the previous review's blocker remains unaddressed: env: on the Run agent step (.github/workflows/pullfrog.yml:87) has only comment children, so it parses to YAML null and GitHub rejects the workflow (Invalid workflow file … Unexpected value ''), which fails every auto-review dispatch before the agent starts. The existing thread carries the fix; do not merge until the env: key is deleted.

Reviewed changes

• Trimmed the agent job's permissions to id-token: write + contents: read — dropped issues: write / pull-requests: write; verified against Pullfrog's documented contract (docs.pullfrog.com/headless-action, "Required permissions") that the action mints its own App installation tokens via OIDC and needs nothing further from GITHUB_TOKEN.
• Rewrote the header credential note — now names pullfrog[bot] as the reviewer identity and records that PR-internal auto-review runs execute master's copy of the workflow, so the identity change is only verifiable post-merge.

  ^{｜ Fix it ➔ ｜ View workflow run ｜ Using DeepSeek Flash ｜ 𝕏}

[eatmoreduck]

blocker 已修（2fec60f）：env: 键整个删除（纯注释子节点解析为 null 会被 GitHub 拒载，评审指得对——且 PR 内 CI 跑的是 master 上的旧定义，绿不证明新 yml 能载入），GH_TOKEN 禁设说明上移为步骤级注释并补记了这个坑。等本 PR 合并后按描述里的流程开测试 PR 验证署名。

[github-actions]

✅ No new issues found.

Reviewed changes

• Deleted the env: key from the Run agent step — the previous review's blocker is fixed: a comment-only env: parsed to YAML null and invalidated the whole workflow. The comments are lifted to step level and a note warns against re-creating the pattern. Verified with yaml.safe_load (no env key) and actionlint 1.7.7 (exit 0, no findings).

  ^{｜ View workflow run ｜ Using DeepSeek Flash ｜ 𝕏}