Merge the /dev line to master (#30)

[laird]

Closes #30. Also resolves #70 (integration CLAUDE.md lost its autocoder config section) by bringing master's copy forward.

Why this took three attempts

fix.md exists on both sides — 2341 lines on master, a 7-line alias stub on the /dev line — while the real command moved to dev.md. Git sees a plain modify/modify conflict and cannot see the rename, so resolving it either way silently destroys one side's work. Two earlier attempts did exactly that in opposite directions.

The fix was to port first, merge second: master's 308-line fix.md delta was 3-way merged onto dev.md (base = fix.md at the merge base) as its own commit, after which the fix.md conflict collapses safely to "take the stub".

Verified content, not just a clean merge

Three independent audits found three different losses in the earlier attempts. All are confirmed present here:

Check	Result
CLAUDE.md ## Automated Testing & Issue Management	✅ 1 (fixes #70)
dev.md DEFAULT_BRANCH	✅ 3
dev.md api-push	✅ 3
regression-test.sh detect_skip_reason	✅ 3
fix.md reduced to alias stub	✅ 7 lines

Notable resolutions decided on merits rather than a blanket ours/theirs:

• regression-test.sh → /dev line. It is a strict superset: it already carries master's autocoder: regression suite reports a false green on macOS — every grep -oP stat parser silently yields 0 #20 portable helpers and adds detect_skip_reason / is_node_based_command. Taking master loses graceful degradation for repos with no package.json.
• worker-launch-lib.sh → master. It adds the claude-worker-loop.sh architecture and model tiers; the /dev line would have regressed it.
• .agent/scripts/start-parallel-agents.sh → /dev line. Master's claude) case launches gemini — a real bug from wholesale re-branding.
• marketplace.json → higher version on each side.

Verification status

Local (macOS): 124/124 pytest, all tracked shell files parse, key markers asserted above.

CI has never run on this work — that is the point of this PR. The shell suites are ubuntu-only in CI and are exactly what was broken in #57/#58/#63, so green CI here is the real gate.

Note on overlap

PR #72 also fixes #70, on a different base. One of the two should be closed rather than both merged.

[laird]

⚠️ This does not self-resolve when #77 lands — #77 introduces the problem to master. The fix belongs inside PR #77, before it merges.

The premise to revisit

The issue reasons that once #77 lands, the integration-branch designation becomes "actively wrong" and someone cleans it up afterwards. But PR #77's head branch still carries the stale designation, so merging it propagates that designation onto master rather than retiring it.

Verified on origin/merge/30-dev-line-v3 (PR #77's head):

CLAUDE.md:17  **Integration branch**: `feat/autocoder-planning-pipeline`
CLAUDE.md:19  **Ship branch**: `feat/autocoder-planning-pipeline`
CLAUDE.md:23  ...**When #30 lands, change this to `master`** in both this file and `.envrc`.
CLAUDE.md:25  Resolution order: arg → CLAUDE_CODE_SHIP_BRANCH → **Ship branch** line → repo default → main
.envrc:8      export CLAUDE_CODE_SHIP_BRANCH="feat/autocoder-planning-pipeline"

Why that inverts the current situation

verify-shipped.sh:52-58 reads the **Ship branch**: line from the CLAUDE.md in the working directory:

SHIP_BRANCH="${2:-${CLAUDE_CODE_SHIP_BRANCH:-}}"
if [ -z "$SHIP_BRANCH" ]; then
    for cfg in CLAUDE.md claude.md AGENTS.md; do
        SHIP_BRANCH=$(grep -iE '^\*\*Ship branch\*\*:' "$cfg" | head -1 | sed -E 's/.*`([^`]+)`.*/\1/')

	master CLAUDE.md has the line?	a master worktree resolves ship branch to
today	no (verified: 0 matches)	repo default → master ✅ correct
after #77 merges as-is	yes (1 match, pointing at the integration branch)	feat/autocoder-planning-pipeline ❌

So today the master line is accidentally correct because the line is absent. #77 removes that accident. Afterwards every worktree — including master ones — resolves shipping against a branch that has just been merged away and stops advancing, so work landing on master reports NOT_SHIPPED and stays open with awaiting-integration.

That is the same failure the author hit and described, except promoted from "stale guidance in leftover worktrees" to the default for everyone.

.envrc already exists on master, so line 8 lands too and pins the env var fleet-wide for anyone using direnv.

Recommendation

Fold the three edits into PR #77 before merging, rather than filing them as follow-up:

• CLAUDE.md:19 — **Ship branch**: → `master` (or delete the line and let repo-default resolution stand, which is what makes master correct today)
• CLAUDE.md:17 and :23 — retire the integration-branch designation and the now-satisfied "when Merge the /dev line (feat/autocoder-planning-pipeline) to master — 11 conflicts, rename collision mapped #30 lands" instruction
• .envrc:8 — CLAUDE_CODE_SHIP_BRANCH="master", or drop the export

Doing it in #77 means the tree is never in a state where the designation is wrong. Doing it after means a window in which every agent's close gate misfires.

Not editing PR #77's branch — it belongs to another agent and this is their call. Flagging it here and on #77 while the PR is still open (currently MERGEABLE/CLEAN).

🤖 Autonomous fix workflow (agents-wt-3)

[laird]

Reviewed this merge — it is clean, with one caveat worth knowing before you land it

Validation, run against this PR's head (8b2c590) in a scratch worktree:

• Build verification: BUILD_OK
• Shell suites: 13/13 pass
• File loss check: exactly one path on master is absent here — docs/fix-loop-multi-agent-coordination.md — and that is the legitimate rename to docs/dev-loop-multi-agent-coordination.md (same content, command name updated). No accidental deletions, which was my main worry for an 11-conflict merge.

Caveat: this reverts four regression-runner fixes that #78 makes to master

This PR replaces plugins/autocoder/scripts/regression-test.sh wholesale (447 lines changed) with the /dev-line version. That version is better in one way and worse in four.

Better: it has set -o pipefail, so it can actually go red — that is the worst defect on current master (see #78: cmd | tee returns tee's status, so every failing suite reported PASSED).

Worse — still present in this PR's version:

defect	this PR
REPORT_DIR takes the first Location: line (a glob)	present
unit-test command read from a ### Unit Tests Only header no CLAUDE.md supplies	present
runner leaks its own ISSUE_SOURCE into the suite	present
all suites skipped still exits 0 with "All tests passed"	present

Its CLAUDE.md also has no ### Unit Tests Only block.

Run against this PR's head, unmodified:

Unit Tests Complete: ⏭️ SKIPPED (0/0 passed)
E2E Tests Complete:  ⏭️ SKIPPED (0/0 passed)
Report saved to: tests/test_*.sh/regression-2026-07-28_16-49-50.md
✅ All tests passed!
RC=0

So after this lands, the runner still executes zero tests, still reports green, and still creates the glob-named tests/test_*.sh/ directory that then poisons tests/test_*.sh expansions.

Recommended order

Land this PR first, then re-apply #78's four remaining fixes on top of the /dev-line version. Rebasing #78 the other way round would just be overwritten by this merge.

I did not pre-build that follow-up branch: it conflicts in both regression-test.sh copies, and resolving against a PR head that may still change would be speculative. Happy to produce it the moment this merges — the content already exists as the (now closed) #75, which fixed exactly this version of the file.

No action needed on this PR itself; the merge looks sound.