Repository navigation
feat: DOTBOT_ACT_MODEL for the act model, plus the fixes that unblock dotbot approval - #7
Merged
Merged
Conversation
Review already takes its model from the DOTBOT_REVIEW_MODELS variable (first entry = primary reviewer, the rest = the "fight" roster), rendered into .openrouter-review.yml for the run. Act had no equivalent: the act model could only be changed by editing act.model in the consuming repo's committed config. Add the act counterpart: * dotbot-act.yml renders `act:\n model: <slug>` from vars.DOTBOT_ACT_MODEL and passes it through the `config_path` action input. The variable is a REPLACEMENT, not an addition - when set it is authoritative, when unset the in-repo file (or action default) stands. * The generated file is written to $RUNNER_TEMP, NOT the checkout: act mode pushes commits to the PR branch, so a generated file in the worktree could be swept into the agent's commit and overwrite the repo's own config. Review can render in place because review never commits. * Slugs are validated (`[A-Za-z0-9._:/-]+`) so a malformed variable fails the run instead of producing a broken config file. Tests: workflow wiring assertions for both variables, plus model-config tests proving an act-only generated file outside GITHUB_WORKSPACE is loaded and that it wins over the repo's committed act.model. Docs: new "Org/Repo Variables (vars.DOTBOT_*)" README section covering both variables, their precedence, and why act renders outside the checkout.
The dotCMS org variable carries ~deepseek/deepseek-flash-latest, and the `~` "latest" alias prefix is a normal part of OpenRouter slugs — the first revision of the guard rejected it and would have failed every act run. * Widen the slug guard to [A-Za-z0-9._:/~@+-]. * Emit the rendered value as a single-quoted YAML scalar, so unusual-but-legal slugs survive the round trip. * Add tests/test_dotbot_workflow_model_vars.py: executes the workflow step's real `run:` script under bash and asserts the generated config (parsed back through load_model_config) for ~latest aliases, :free variants, whitespace stripping, the blank-variable fallback, and the non-slug failure path. Assertions on the script's internals in test_module_coverage.py are dropped in favour of this.
Member
Author
|
Follow-up commit
Verified against the current org values:
|
The self-review workflow passed no `github_approval_token`, so every run ended with "[debug1] All reviewers agree the patch is correct; no approval token configured" and dotbot never approved: PRs needing 1 approving review stalled. This repo has no DOTBOT_GITHUB_USER_PAT (0 repo secrets; the org secret does not exist either), and REPO_ACCESS_TOKEN belongs to the PR author, whose approvals GitHub rejects — which is why the earlier machine-user wiring was dropped in a26dc15. The repo does allow Actions to approve PRs (can_approve_pull_request_reviews: true), so pass `github.token` instead: the approval lands as github-actions[bot] and satisfies the ruleset's 1-approval / last-push-approval requirements. Tradeoff documented in the workflow comment and README: a bot identity cannot approve a PR the bot authored, and a machine-user PAT remains the better option where one exists.
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
Pinning to d68edbb (the split-review-act era) meant GitHub silently rejected inputs that revision did not declare: ##[warning]Unexpected input(s) 'github_approval_token', valid inputs are [...] So `github_approval_token: ${{ github.token }}` never reached the action and auto-approval still did not fire after the previous commit. Bump both workflows to bbe2345 (v1.2.0 / v1 / latest), which declares the inputs the workflows pass. Also assert in tests that both workflows pin the same released commit — drift between them is how the review side ended up on a pre-approval revision.
wezell/openrouter-code-review-action is no longer the source of releases: its main/v1/latest sit at 34ee169 (v1.0.0-era, no github_approval_token input), and it has no release carrying that input. So after the previous two commits the run died earlier still: ##[error]Unable to resolve action `wezell/openrouter-code-review-action@bbe2345...` This repo is canonical and cuts its own releases (auto-release.yml), so point both workflows at bbe2345... (v1.2.0 / v1 / latest), which declares every input the workflows pass, and require the canonical owner in tests. Supersedes the still-open PR #4, which pinned the review workflow to v1.1.0 and left act on wezell.
With the input now wired, the run got as far as:
Warning: failed to submit PR approval: failed to resolve approval token user:
Resource not accessible by integration: 403
`submit_pr_approval` resolved the token's login via `GET /user` before doing
anything else. Installation tokens (`${{ github.token }}`) cannot call that
endpoint, so the approval was abandoned — the input reached the action and was
still useless.
- `_resolve_login` is now best-effort: an unresolvable token logs a debug line
and continues with an empty login instead of raising.
- Idempotency is keyed off APPROVAL_MARKER ("approved automatically by dotbot")
in the review body in addition to the login, so re-runs skip duplicate
approvals even when the approver identity cannot be read. Extracted the
APPROVAL_MARKER constant so the body template and the check cannot drift.
- Failure to load the PR still raises (unchanged); the approval POST failure
message names the token as an installation token when the login is unknown,
and the workflow's log lines omit "as <login>" rather than printing an empty
name.
Also documents the choice in action.yml, README, and the `_maybe_approve_pr`
docstring.
Composite `run` steps execute in the caller's workspace, and `python -m` puts the current directory ahead of PYTHONPATH on sys.path. So `python3 -u -m cli.main` loaded `cli/` from the *checkout* whenever the consuming repo shipped one — for this repo's own self-review, the PR head's code ran with the workflow token instead of the pinned action. Demonstrated on PR #7: the run reported "Approved PR #7: all reviewers agree the patch is correct" (new message format from that very PR), and it submitted an approval the pinned revision cannot submit (its approval.py aborts on GET /user with 403). For the act workflow this is the escalation the trusted-ref pin exists to prevent: a /dotbot comment on an untrusted PR head would execute that PR's `cli/` package with `contents: write` and REPO_ACCESS_TOKEN. Set PYTHONSAFEPATH=1 on both CLI-invoking steps (3.11+: do not prepend cwd to sys.path), and add a test that every CLI step in action.yml carries it. Repro: mkdir -p action/cli work/cli; ... # both define cli.main cd work && PYTHONPATH=…/action python3 -m cli.main # -> ran from: work cd work && PYTHONSAFEPATH=1 PYTHONPATH=…/action python3 -m cli.main # -> action
|
dotbot code review:
PYTHONSAFEPATH on both CLI-invoking composite steps correctly prevents checkout cli/ shadowing while preserving PYTHONPATH lookup; coverage test enforces the invariant. Tip: comment with "/dotbot address comments" to attempt automated fixes for unresolved review threads. reviewed by dotbot · meta/muse-spark-1.3 · medium |
|
dotbot code review:
PYTHONSAFEPATH=1 is set on both CLI-invoking steps (prepare_resume_state and cli.main), correctly preventing Tip: comment with "/dotbot address comments" to attempt automated fixes for unresolved review threads. reviewed by dotbot · ~z-ai/glm-latest · medium |
wezell
changed the title
This was referenced Sep 29, 2026
wezell
added a commit
that referenced
this pull request
Sep 29, 2026
v1.3.1 (3f4cfa1) is the first release whose action.yml loads: v1.3.0 shipped a `${{ github.token }}` in an input description, which made GitHub reject the whole action, so pinning 1.3.0 left this repo's own review/act jobs failing at "Set up job" (that is what this PR's earlier run hit). v1.3.1 carries everything from #7 plus the description fix, so the self-review now runs the released action instead of executing the checkout's cli/ package. v1/latest point at the same commit; the test constant moves with the pins.
wezell
added a commit
that referenced
this pull request
Sep 29, 2026
v1.3.1 (3f4cfa1) is the first release whose action.yml loads: v1.3.0 shipped a `${{ github.token }}` in an input description, which made GitHub reject the whole action, so pinning v1.3.0 left this repo's own review/act jobs failing at "Set up job" (that is what this PR's earlier run hit). v1.3.1 carries everything from #7 plus the description fix, so the self-review now runs the released action instead of executing the checkout's cli/ package. v1/latest point at the same commit; the test constant moves with the pins.
This was referenced Sep 29, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What / why
Started as "act model should come from an org/repo variable like the review model already does", and turned into a chain of fixes uncovered while trying to get this very PR approved. 6 commits.
1.
DOTBOT_ACT_MODELdrives the act model (207b094,4fadf9f)Review already took its model from
DOTBOT_REVIEW_MODELS(first entry = primary reviewer, the rest = the "fight" roster); act had no equivalent. Now:DOTBOT_REVIEW_MODELSDOTBOT_ACT_MODEL/dotbotedits.openrouter-review.yml/ action default.$RUNNER_TEMP/dotbot-act-model.ymland passed viaconfig_path, never written into the checkout — act mode pushes to the PR branch, so a generated file in the worktree could be swept into the agent's commit.~-prefixed "latest" aliases (the org value is~deepseek/deepseek-flash-latest; the first draft rejected them and would have failed every act run) and:free-style variants.2. Auto-approval was never wired (
1427f3d)Every review run ended with
[debug1] All reviewers agree the patch is correct; no approval token configured: the workflow passed nogithub_approval_token, and noDOTBOT_GITHUB_USER_PATexists (0 repo secrets; the org secret does not exist either).REPO_ACCESS_TOKENbelongs to the PR author, whose approvals GitHub rejects — why the machine-user wiring was dropped ina26dc15. Passinggithub.tokenis what the repo permits (can_approve_pull_request_reviews: true).3. The pin silently dropped the input (
3665e57,863820f)wezell/openrouter-code-review-action@d68edbbpredates thegithub_approval_tokeninput, so GitHub discarded it:##[warning]Unexpected input(s) 'github_approval_token'.wezell/…is no longer the release source (itsv1/latest=34ee169, v1.0.0-era). Both workflows now pin the canonicaldotCMS/openrouter-code-review-action@bbe2345(v1.2.0 / v1 / latest), and a test asserts both workflows pin the same released SHA. Supersedes #4.4. Installation tokens could not submit the approval (
b6ed738)submit_pr_approvalresolved the token's login viaGET /userfirst; installation tokens cannot call it. Login resolution is now best-effort and idempotency is keyed offAPPROVAL_MARKERin the review body as well, so re-runs still skip duplicates when the approver identity is unreadable.5. P0: the checkout shadowed the pinned action (
5bd98e4)Composite
runsteps execute in the caller's workspace andpython -mputs the cwd ahead ofPYTHONPATH, sopython3 -u -m cli.mainloadedcli/from the checkout whenever the consuming repo ships one. Evidence from this PR: the run atb6ed738printedApproved PR #7: all reviewers agree the patch is correct— the new message format from that very commit — and submitted an approval the pinned revision cannot submit.For
dotbot-act.ymlthat is the exact escalation the trusted-ref pin exists to prevent: a/dotbotcomment on an untrusted PR head would run that PR'scli/package withcontents: writeandREPO_ACCESS_TOKEN. Fix:PYTHONSAFEPATH=1on both CLI-invoking steps, plus a test that every CLI step inaction.ymlcarries it.Verification
uv run pytest -q→ 806 passed;pre-commit run --all-files(ruff format/check, mypy) → Passedtests/test_dotbot_workflow_model_vars.pyexecutes the act step's realrun:script under bash and parses the generated config back throughload_model_configgithub-actions APPROVEDat the head SHAMerge status
Merge-ready: all checks pass,
reviewDecision: APPROVED,mergeStateStatus: CLEAN, and every review thread resolved.Correction to an earlier note in this PR: the ruleset block was not the approval identity —
github-actions[bot]'s approval does count (reviewDecisionflipped toAPPROVEDwith it). The PR was blocked byrequired_review_thread_resolution: the two unaddressed dotbot findings were still open. Once answered and resolved, the PR went straight toCLEAN. Worth remembering that dotbot threads have to be resolved, not just fixed.After merge,
auto-release.ymlcutsv1.3.0and retagsv1/latest; a follow-up bumps the self-review pins to that SHA so thePYTHONSAFEPATHand approval fixes apply to the self-review itself (until then this PR's own runs still execute the checkout's CLI — see commit 5).