ajsai47/holyclaude-cloud/reconciler

legion-reconciler

Dependency-ordered PR merging. Mediator-aware.

Flow per invocation

One pass:

1. legion reconcile finds tasks whose status == shipped, have a pr_url, have no merge_blocker, and whose deps are all merged.
2. For each:

• CI check (gh pr checks <pr-number>):
• pending → skip this tick, try again next pass
• fail → mark merge_blocker = "ci_failed", surface to user
• pass or none → proceed
• Merge attempt (gh pr merge --squash --delete-branch):
• merged → mark merged_at = now()
• conflict → invoke mediator (see below)
• gh_error → mark merge_blocker = "gh_error", surface

1. Emit a JSON list of actions taken.

The reconciler is idempotent. You can (and should) call legion reconcile repeatedly while there are shipped-but-unmerged tasks. In the orchestrator's main loop, interleave legion poll (worker status) with legion reconcile (PR status).

Mediator fallback

On merge conflict:

1. Increment task.mediator_attempts.
2. If attempts > config.reconciler.mediator_max_retries (default 2), mark merge_blocker = "mediator_maxed" and stop.
3. Otherwise call mediator.run_mediator(task, base_branch):

• Creates a fresh worktree at .legion/mediators/<task-id>/ off origin/<base>.
• git merge origin/<task.branch> — forces conflict markers to appear.
• Spawns claude -p with a mediator-framed prompt (preserve both intents, don't revert main's commits, don't touch unrelated files).
• On clean exit + no remaining <<<<<<< markers: commit + force-push to task.branch.
• Retry gh pr merge once.

If the mediator resolves the conflict and the retry merge succeeds: task status becomes mediated_and_merged.

If the mediator leaves conflicts or fails: mediation_failed with detail, task gets a blocker, surface to user.

When to invoke from a skill

• After orchestrator's dispatch loop finishes — shipped tasks accumulate; drain them with reconcile.
• Interleaved inside the dispatch loop — as workers finish, merge them immediately so downstream deps can start.
• Manually — /legion-reconcile for a one-shot, legion mediate <task-id> to re-run the mediator on a specific task.

Failure surfaces

• merge_blocker = "ci_failed": user's CI is telling the task it's broken. Orchestrator should NOT re-dispatch automatically — the code is wrong, not the run. Surface to user with the PR URL + check detail.
• merge_blocker = "mediator_maxed": the mediator tried mediator_max_retries times and couldn't resolve. Surface the PR + the mediator log (.legion/mediator_logs/<task-id>.log) and ask the user to resolve manually.
• merge_blocker = "needs_human": reserved for Phase 4 — user explicitly escalated.
• merge_blocker = "gh_error": gh CLI returned an error we don't know how to classify. Usually means auth, rate limit, or branch state issues. Show the raw error.

Subcommands used

legion reconcile                # one idempotent pass
legion mediate <task-id>        # force mediator run on a task
legion status                   # shows merged/unmerged/blocked