[B-REL] release executor: idempotent half-landed retry + commit-scoped CI + decoupled workflow

Three confirmed gaps in the release fail-closed pipeline (#87/#318/#402):

#87 publish_failed retry duplicates changelog: execute() only short-circuits
on an existing tag. A publish_failed outcome (commit pushed + CI green, no
tag) left no tag, so a retry re-ran apply_version_bumps + write_changelog_entry
(re-inserting the entry above the already-present heading -> duplicate) and
commit_and_push (a second chore(release) commit). Add ReleaseOps
.release_commit_sha(version) detecting a prior release commit on the branch
(clone already at the target version); when present, skip the bump/changelog/
gate/commit pipeline and rejoin the shared CI -> publish tail on the existing
commit. No second commit, no duplicate entry.

#318 wait_for_ci polls branch-latest, not the release commit: a later push to
master during the ~40min wait made the latest run's head_sha != the release
sha forever, exhausting _CI_MAX_POLLS -> false ci_failed on a release whose
own CI was green. Thread head_sha through get_latest_ci_conclusion /
_fetch_latest_ci_run (GitHub actions/runs?head_sha=) so the gate polls the
release commit's own run; a concurrent push can no longer mask it.

#402 release CI gate reuses self_heal_ci_workflow: that setting documents an
empty-string mode for single-workflow repos which, inherited here, degraded
the fail-closed gate to the all-workflows mode git.py itself flags as
unreliable. Add release_ci_workflow (default ci.yml) and _resolve_release_
ci_workflow(); the release gate always resolves a NAMED workflow, never None.

Refactor: bundle the CI-fetch per-project inputs into a _CiRunQuery dataclass
so _fetch_latest_ci_run stays under the arg-count gate; unify the half-landed
path into execute's shared tail (drops a separate _publish_existing, one
return path). TDD red->green; ruff/mypy clean.

roboco/config.py

@@ -594,6 +594,19 @@ class Settings(BaseSettings):
         ge=60,
         description="Seconds between release-readiness assessment passes.",
     )
+    release_ci_workflow: str = Field(
+        default="ci.yml",
+        description=(
+            "GitHub Actions workflow file name the release fail-closed CI gate "
+            "scopes to. Decoupled from self_heal_ci_workflow — that setting "
+            "documents an empty-string mode for single-workflow repos which, "
+            "inherited here, would degrade the release gate to the "
+            "all-workflows mode git.py itself flags as unreliable (a green "
+            "secondary workflow masking a red primary CI). The release gate "
+            "always resolves a NAMED workflow; empty falls back to 'ci.yml', "
+            "never None."
+        ),
+    )
 
     # Organizational-memory loop — distill a high-signal lesson at task
     # completion, index journal reflections, and auto-inject similar past

roboco/services/git.py

@@ -237,6 +237,18 @@ def _select_ci_head_run(runs: list[dict[str, Any]]) -> dict[str, Any]:
     return max(same_head, key=lambda r: int(r.get("run_attempt") or 0))
 
 
+@dataclass(frozen=True)
+class _CiRunQuery:
+    """Bundle of per-project inputs to a CI-run fetch (owner/repo, branch, token,
+    slug for logging) so ``_fetch_latest_ci_run`` stays under the arg-count gate —
+    owner_repo alone was already bundled for the same reason."""
+
+    project_slug: str
+    owner_repo: tuple[str, str]
+    branch: str
+    git_token: str
+
+
 class GitService(BaseService):
     """
     Service for git operations on agent workspaces.
@@ -1933,7 +1945,11 @@ class GitService(BaseService):
         }
 
     async def get_latest_ci_conclusion(
-        self, project_slug: str, *, workflow: str | None = None
+        self,
+        project_slug: str,
+        *,
+        workflow: str | None = None,
+        head_sha: str | None = None,
     ) -> dict[str, Any] | None:
         """Latest completed CI (GitHub Actions) run on a project's default branch.
 
@@ -1944,10 +1960,12 @@ class GitService(BaseService):
         Resolves owner/repo and the git token PER PROJECT. ``workflow`` (a
         workflow file name like ``ci.yml``) scopes the signal to one workflow —
         without it the latest run across ALL workflows is used, which is
-        imprecise on a multi-workflow repo. Returns ``None`` on a missing token,
-        unparseable remote, GitHub error, or a repo with no matching Actions runs
-        (a repo that doesn't use GitHub Actions yields no signal, not a false
-        one). It never raises into the poll loop.
+        imprecise on a multi-workflow repo. ``head_sha`` further scopes the run
+        window to a specific commit (the release gate uses this so a later push
+        to the default branch can't mask the release commit's own CI). Returns
+        ``None`` on a missing token, unparseable remote, GitHub error, or a repo
+        with no matching Actions runs (a repo that doesn't use GitHub Actions
+        yields no signal, not a false one). It never raises into the poll loop.
         """
         project = await get_project_service(self.session).get_by_slug(project_slug)
         if project is None or not project.git_url:
@@ -1960,9 +1978,13 @@ class GitService(BaseService):
         if not git_token:
             return None
         branch = project.default_branch or "master"
-        run = await self._fetch_latest_ci_run(
-            project_slug, (owner, repo), branch, git_token, workflow
+        query = _CiRunQuery(
+            project_slug=project_slug,
+            owner_repo=(owner, repo),
+            branch=branch,
+            git_token=git_token,
         )
+        run = await self._fetch_latest_ci_run(query, workflow, head_sha)
         if run is None:
             return None
         return {
@@ -2013,42 +2035,47 @@ class GitService(BaseService):
 
     async def _fetch_latest_ci_run(
         self,
-        project_slug: str,
-        owner_repo: tuple[str, str],
-        branch: str,
-        git_token: str,
+        query: _CiRunQuery,
         workflow: str | None = None,
+        head_sha: str | None = None,
     ) -> dict[str, Any] | None:
         """Resolve ``branch``'s current-HEAD CI conclusion; None on error.
 
         Scopes to ``workflow`` (a workflow file name) when given — the precise
         signal — otherwise reads across ALL workflows, which on a multi-workflow
-        repo is unreliable (an unrelated green run can mask a red CI run). Pulls a
-        WINDOW of recent completed runs and selects the newest commit's latest
-        attempt (see ``_select_ci_head_run``) rather than the single
-        most-recently-completed run, so a green run on an older commit can't mask
-        the HEAD's failure and a green re-run correctly supersedes it. ``branch``
-        filters by head branch, so only pushes to the default branch (not
-        pull-request runs, whose head is a feature branch) count — exactly the
-        "is the default branch red" signal self-heal needs. Transient network /
-        429 / 5xx errors are retried a few times before giving up so a single
-        blip doesn't silently skip the cycle.
+        repo is unreliable (an unrelated green run can mask a red CI run).
+        ``head_sha`` further scopes the window to one commit so the release gate
+        can wait on a SPECIFIC release commit's CI without a later push to the
+        default branch masking it. Pulls a WINDOW of recent completed runs and
+        selects the newest commit's latest attempt (see
+        ``_select_ci_head_run``) rather than the single most-recently-completed
+        run, so a green run on an older commit can't mask the HEAD's failure and
+        a green re-run correctly supersedes it. ``branch`` filters by head
+        branch, so only pushes to the default branch (not pull-request runs,
+        whose head is a feature branch) count — exactly the "is the default
+        branch red" signal self-heal needs. Transient network / 429 / 5xx errors
+        are retried a few times before giving up so a single blip doesn't
+        silently skip the cycle.
         """
-        owner, repo = owner_repo
+        owner, repo = query.owner_repo
         api_base = settings.github_api_base_url.rstrip("/")
         base = f"{api_base}/repos/{owner}/{repo}/actions"
         url = f"{base}/workflows/{workflow}/runs" if workflow else f"{base}/runs"
         headers = {
-            "Authorization": f"Bearer {git_token}",
+            "Authorization": f"Bearer {query.git_token}",
             "Accept": "application/vnd.github+json",
             "X-GitHub-Api-Version": "2022-11-28",
         }
         params: dict[str, str | int] = {
-            "branch": branch,
+            "branch": query.branch,
             "status": "completed",
             "per_page": _CI_RUN_WINDOW,
         }
-        resp = await self._get_ci_runs_response(project_slug, url, headers, params)
+        if head_sha:
+            params["head_sha"] = head_sha
+        resp = await self._get_ci_runs_response(
+            query.project_slug, url, headers, params
+        )
         if resp is None or not resp.is_success:
             return None
         data = resp.json()

roboco/services/release_executor.py

@@ -50,6 +50,8 @@ class ReleaseOps(Protocol):
 
     async def is_already_published(self, version: str) -> bool: ...
 
+    async def release_commit_sha(self, version: str) -> str | None: ...
+
     async def apply_version_bumps(
         self, plan: list[str], new_version: str
     ) -> list[str]: ...
@@ -72,7 +74,14 @@ class ReleaseExecutor:
         self._ops = ops
 
     async def execute(self, report: ReleaseReadinessReport) -> ReleaseResult:
-        """Bump → gate → commit/push → CI → publish, aborting on any red step."""
+        """Bump → gate → commit/push → CI → publish, aborting on any red step.
+
+        A half-landed (publish_failed) retry — a prior ``chore(release): {ver}``
+        commit already on the branch but no tag — skips the bump/changelog/gate/
+        commit pipeline (which would duplicate the changelog entry and land a
+        second release commit) and rejoins the shared CI → publish tail on the
+        existing commit.
+        """
         version = report.proposed_version
         if await self._ops.is_already_published(version):
             return ReleaseResult(
@@ -84,38 +93,52 @@ class ReleaseExecutor:
                 detail=f"v{version} is already published; nothing to do.",
             )
 
-        files = await self._ops.apply_version_bumps(report.version_bump_plan, version)
-        await self._ops.write_changelog_entry(report.drafted_changelog)
-
-        if not await self._ops.run_gate():
-            return ReleaseResult(
-                status="gate_failed",
-                version=version,
-                files_changed=files,
-                commit_sha=None,
-                release_url=None,
-                detail="make quality failed — aborted before commit (fail-closed).",
+        # Half-landed detection (publish_failed retry): a prior release commit
+        # on the branch means the pipeline already ran — only the publish step
+        # failed. Re-running it would re-insert the changelog entry (duplicate,
+        # above the already-present heading) and land a SECOND release commit.
+        # Reuse the existing commit and rejoin the shared CI → publish tail.
+        # None when no prior release commit exists (fresh release).
+        existing_sha = await self._ops.release_commit_sha(version)
+        if existing_sha is not None:
+            commit_sha = existing_sha
+            files: list[str] = []
+        else:
+            files = await self._ops.apply_version_bumps(
+                report.version_bump_plan, version
             )
+            await self._ops.write_changelog_entry(report.drafted_changelog)
 
-        try:
-            commit_sha = await self._ops.commit_and_push(version)
-        except RuntimeError as exc:
-            # The ops layer raises RuntimeError on a failed add/commit/push
-            # (gpgsign/pre-commit reject/no-op bump/non-fast-forward push). That
-            # is the correct fail-closed abort at the ops layer; the EXECUTOR
-            # turns it into a structured outcome so the CEO sees the cause
-            # instead of a 500 bubbling out of ``approve``.
-            logger.error("release commit/push failed", error=str(exc)[:300])
-            return ReleaseResult(
-                status="commit_failed",
-                version=version,
-                files_changed=files,
-                commit_sha=None,
-                release_url=None,
-                detail=(
+            if not await self._ops.run_gate():
+                return ReleaseResult(
+                    status="gate_failed",
+                    version=version,
+                    files_changed=files,
+                    commit_sha=None,
+                    release_url=None,
+                    detail="make quality failed — aborted before commit (fail-closed).",
+                )
+
+            try:
+                commit_sha = await self._ops.commit_and_push(version)
+            except RuntimeError as exc:
+                # The ops layer raises RuntimeError on a failed add/commit/push
+                # (gpgsign/pre-commit reject/no-op bump/non-fast-forward push).
+                # That is the correct fail-closed abort at the ops layer; the
+                # EXECUTOR turns it into a structured outcome so the CEO sees
+                # the cause instead of a 500 bubbling out of ``approve``.
+                logger.error("release commit/push failed", error=str(exc)[:300])
+                detail = (
                     f"release commit/push failed — not published (fail-closed): {exc}"
-                )[:280],
-            )
+                )[:280]
+                return ReleaseResult(
+                    status="commit_failed",
+                    version=version,
+                    files_changed=files,
+                    commit_sha=None,
+                    release_url=None,
+                    detail=detail,
+                )
 
         if not await self._ops.wait_for_ci(commit_sha):
             return ReleaseResult(
@@ -238,6 +261,29 @@ class _GitReleaseOps:
         rc, out = await self._git("ls-remote", "--tags", "origin", f"v{version}")
         return rc == 0 and bool(out.strip())
 
+    async def release_commit_sha(self, version: str) -> str | None:
+        """Detect a half-landed release: a ``chore(release): {version}`` commit
+        already on the branch (a publish_failed retry — commit pushed, no tag
+        yet). Returns that commit's sha, or None when the clone is not yet at
+        ``version`` (no bump happened) or no such commit exists.
+
+        The clone is fresh per execute(), so if the working version already
+        equals ``version`` AND the release commit is in history, the pipeline
+        ran in a prior attempt and must not re-run (it would duplicate the
+        changelog entry and land a second release commit).
+        """
+        if self._current_version() != version:
+            return None
+        rc, out = await self._git("log", "-n", "50", "--format=%H%x00%s")
+        if rc != 0:
+            return None
+        wanted = f"chore(release): {version}"
+        for line in out.splitlines():
+            sha, _, subject = line.partition("\x00")
+            if sha and subject.strip() == wanted:
+                return sha.strip()
+        return None
+
     def _current_version(self) -> str:
         text = (self._root / "pyproject.toml").read_text(encoding="utf-8")
         match = re.search(r'^version\s*=\s*"([^"]+)"', text, re.MULTILINE)
@@ -316,7 +362,7 @@ class _GitReleaseOps:
         git = get_git_service(self._session)
         for _ in range(_CI_MAX_POLLS):
             ci = await git.get_latest_ci_conclusion(
-                self._slug, workflow=self._ci_workflow
+                self._slug, workflow=self._ci_workflow, head_sha=commit_sha
             )
             if ci and ci.get("head_sha") == commit_sha:
                 conclusion = (ci.get("conclusion") or "").lower()
@@ -378,6 +424,20 @@ def _insert_changelog_entry(existing: str, entry: str) -> str:
     return existing.rstrip() + "\n\n" + block
 
 
+def _resolve_release_ci_workflow() -> str:
+    """The release CI gate's workflow, decoupled from self_heal_ci_workflow.
+
+    ``self_heal_ci_workflow`` documents an empty-string mode for single-workflow
+    repos; inheriting that here would degrade the release fail-closed gate to
+    the all-workflows mode ``_fetch_latest_ci_run`` itself flags as unreliable
+    (a green secondary workflow masking a red primary CI). The release gate
+    always uses a NAMED workflow — empty falls back to ``ci.yml``, never None.
+    """
+    from roboco.config import settings
+
+    return settings.release_ci_workflow or "ci.yml"
+
+
 async def get_release_executor(session: AsyncSession) -> ReleaseExecutor:
     """Build a ReleaseExecutor with a production ops over a fresh writable clone."""
     from roboco.config import settings
@@ -399,7 +459,7 @@ async def get_release_executor(session: AsyncSession) -> ReleaseExecutor:
         default_branch=default_branch,
         root=root,
         auth_url=auth_url,
-        ci_workflow=(settings.self_heal_ci_workflow or None),
+        ci_workflow=_resolve_release_ci_workflow(),
     )
     return ReleaseExecutor(_GitReleaseOps(session, ctx))
 

tests/unit/services/test_release_executor.py

@@ -8,10 +8,25 @@ call sequence; the production git/gh ops is exercised live (CEO-gated).
 
 from __future__ import annotations
 
+from types import SimpleNamespace
+from typing import TYPE_CHECKING
+from unittest.mock import MagicMock
+
 import pytest
-from roboco.services.release_executor import ReleaseExecutor, ReleaseResult
+from roboco.config import settings
+from roboco.services import release_executor as re
+from roboco.services.release_executor import (
+    ReleaseExecutor,
+    ReleaseResult,
+    _GitReleaseOps,
+    _ReleaseContext,
+    _resolve_release_ci_workflow,
+)
 from roboco.services.release_readiness import ReleaseReadinessReport
 
+if TYPE_CHECKING:
+    from pathlib import Path
+
 _PLAN = ["pyproject.toml", "roboco/__init__.py", "CHANGELOG.md"]
 _VERSION = "0.13.0"
 _ONE = 1
@@ -49,14 +64,28 @@ class _FakeOps:
         self._ci = ci
         self._commit_raises = commit_raises
         self._publish_raises = publish_raises
+        # Half-landed (publish_failed retry) detection: a prior
+        # ``chore(release): {version}`` commit already on the branch. Set on the
+        # instance (not via __init__ — keeps the constructor under the arg-count
+        # gate) by tests that exercise the retry path.
+        self._existing_sha: str | None = None
         self.calls: list[str] = []
         self.bumped_plan: list[str] | None = None
         self.bumped_version: str | None = None
+        self.halflanded_check = False
 
     async def is_already_published(self, _version: str) -> bool:
         self.calls.append("check")
         return self._already
 
+    async def release_commit_sha(self, _version: str) -> str | None:
+        # Half-landed detection: a prior `chore(release): {version}` commit
+        # already on the branch means a publish_failed retry must NOT re-run the
+        # bump→changelog→gate→commit pipeline. Recorded via a flag (not calls)
+        # so the green-path call-sequence assertion is unaffected.
+        self.halflanded_check = True
+        return self._existing_sha
+
     async def apply_version_bumps(self, plan: list[str], new_version: str) -> list[str]:
         self.calls.append("bump")
         self.bumped_plan = list(plan)
@@ -186,3 +215,103 @@ def test_release_result_carries_outcome_fields() -> None:
     assert result.version == _VERSION
     assert result.files_changed == _PLAN
     assert result.release_url is not None
+
+
+@pytest.mark.asyncio
+async def test_half_landed_retry_skips_bump_and_republishes_only() -> None:
+    """#87: a publish_failed retry (commit pushed + CI green, no tag yet) must
+    NOT re-run bump/changelog/gate/commit — that would re-insert the changelog
+    entry above the already-present ``## [X.Y.Z]`` heading (duplicate) and land a
+    second ``chore(release): X.Y.Z`` commit. The executor detects the
+    half-landed state via ``release_commit_sha`` (a prior release commit already
+    on the branch) and jumps straight to wait_for_ci + publish."""
+    ops = _FakeOps()
+    ops._existing_sha = "existingbeef"
+    result = await ReleaseExecutor(ops).execute(_report())
+    assert result.status == "published"
+    assert result.commit_sha == "existingbeef"
+    assert result.release_url is not None
+    assert ops.halflanded_check is True
+    assert "bump" not in ops.calls
+    assert "changelog" not in ops.calls
+    assert "gate" not in ops.calls
+    assert "commit" not in ops.calls
+    assert ops.calls == ["check", "ci", "publish"]
+
+
+@pytest.mark.asyncio
+async def test_wait_for_ci_scoped_to_release_commit_not_branch_latest(
+    monkeypatch: pytest.MonkeyPatch, tmp_path: Path
+) -> None:
+    """#318: a later commit landing on master during the ~40min wait must not
+    mask the release commit's green CI. ``wait_for_ci`` scopes the GitHub query
+    to the release commit_sha (``head_sha=``), so the branch-latest run (a later
+    sha) can't make the gate poll forever and false-fail as ci_failed."""
+    commit_sha = "release_commit_abc"
+    later_sha = "later_landed_def"
+
+    async def _fake_get_ci(_slug: str, **_kwargs: object) -> dict[str, str]:
+        # Mimic GitHub's head_sha filter: a run for the release sha only when
+        # asked for it (head_sha=commit_sha); the branch-latest (later commit)
+        # run otherwise. The release gate MUST scope to commit_sha to see green.
+        if _kwargs.get("head_sha") == commit_sha:
+            return {
+                "head_sha": commit_sha,
+                "conclusion": "success",
+                "run_url": "u",
+                "run_name": "n",
+                "branch": "master",
+                "completed_at": "t",
+            }
+        return {
+            "head_sha": later_sha,
+            "conclusion": "success",
+            "run_url": "u2",
+            "run_name": "n2",
+            "branch": "master",
+            "completed_at": "t2",
+        }
+
+    monkeypatch.setattr(
+        "roboco.services.git.get_git_service",
+        lambda _session: SimpleNamespace(get_latest_ci_conclusion=_fake_get_ci),
+    )
+    monkeypatch.setattr(re, "_CI_MAX_POLLS", 2)
+
+    async def _no_sleep(_secs: float) -> None:
+        return None
+
+    monkeypatch.setattr(re.asyncio, "sleep", _no_sleep)
+
+    ctx = _ReleaseContext(
+        slug="roboco-api",
+        default_branch="master",
+        root=tmp_path,
+        auth_url="x",
+        ci_workflow="ci.yml",
+    )
+    ops = _GitReleaseOps(session=MagicMock(), ctx=ctx)
+    ok = await ops.wait_for_ci(commit_sha)
+    assert ok is True
+
+
+def test_release_ci_workflow_decoupled_from_self_heal_setting(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """#402: the release CI gate must not inherit ``self_heal_ci_workflow``'s
+    empty-string tuning (documented valid for single-workflow repos), which would
+    degrade the fail-closed gate to the all-workflows mode git.py itself flags as
+    unreliable. The release gate always resolves a named workflow (default
+    ``ci.yml``), never None."""
+    # The dangerous tuning an operator might apply for self-heal on a
+    # single-workflow repo — must NOT leak into the release gate.
+    monkeypatch.setattr(settings, "self_heal_ci_workflow", "")
+    monkeypatch.setattr(settings, "release_ci_workflow", "ci.yml")
+    assert _resolve_release_ci_workflow() == "ci.yml"
+
+    monkeypatch.setattr(settings, "release_ci_workflow", "release.yml")
+    assert _resolve_release_ci_workflow() == "release.yml"
+
+    # An empty release setting never falls through to None — always the default.
+    monkeypatch.setattr(settings, "release_ci_workflow", "")
+    assert _resolve_release_ci_workflow() == "ci.yml"