> **Based on `main`** — independent of my U7 stack and of #2561; merges
in any order.
Third unowned drift-review site. `isRunnableAutoClaimCandidate` is the
single source of truth for *"may an agent claim this task?"* (FN-6873),
and it carried **three** lifecycle literals that fail in **opposite
directions**.
## The two failures
**`column === "todo"` gated candidacy** on the hold role. Keyed on the
literal, a renamed workflow's candidate set was **permanently empty** —
agents were never offered its work, and nothing anywhere reported it.
Silence, not an error.
**`dependency?.column === "done" || "archived"` gated dependency
satisfaction**, and this is the more dangerous half: a dependency that
finished in a renamed **complete** column was never recognised as done,
so the dependent stayed **blocked forever**.
One makes work invisible; the other makes it permanently ineligible.
Both are silent.
## Roles resolve per task, not per pass
The non-obvious part: **a dependency may sit on a different workflow
from the claimant.** A single per-pass answer is wrong for one of them
on any mixed board — so the map is keyed by task id, and the dependency
check reads the *dependency's* roles, not the claimant's.
Asserted directly: a dependency completed in `done` (default vocabulary)
satisfying a claimant waiting in `drafting` (renamed).
## Shape
Both callers already have the store and are async, so they resolve for
real rather than taking the injected-lane fallback the *synchronous*
predicates needed (#2551). The predicate itself stays synchronous — a
resolved-roles map is passed in — because it runs inside two
`filter`/`flatMap` bodies.
Tasks absent from the map keep the legacy ids, so a partially-resolvable
board degrades to today's behavior instead of silently emptying the
candidate set.
**Type narrowing preserved.** The two callers take `Pick<TaskStore,
"listTasks">`, which is what makes them testable without a real store.
Rather than widening to the whole `TaskStore`, they now take
`Pick<TaskStore, "listTasks"> & WorkflowIrResolverStore` — the minimal
additional shape resolution needs.
## Revert proofs, isolated per literal
| Restored | Result |
|---|---|
| hold literal only | **3 of 6 fail** |
| dependency-completion literals only | **1 of 6 fails** |
The three default-vocabulary cases pass under both. Splitting the proof
matters here: it confirms the two halves are **independently**
load-bearing rather than one masking the other — a single combined
revert would have shown 3 failures and told me nothing about the
dependency half.
## Convergence
Measured on `main`, comment-stripped scan of `column === / !== "todo" |
"triage"` in `packages/*/src` excluding tests:
- this file alone: **103 → 102**
- with #2561: **103 → 100**
The `done` / `archived` literals fixed here sit outside that pattern and
are not counted — same caveat as #2561's gridlock `active` filter. Two
PRs now where the real fix is larger than the metric shows.
## Verification
| Check | Result |
|---|---|
| new suite | 6/6 |
| pre-existing auto-claim suite | 17/17, **no expectation edits** |
| `tsc --noEmit` (engine) | clean |
| `pnpm lint` | clean |
| `pnpm test:gate` | green (414 + 10 + 71) |
| `pnpm check:changesets` | clean |
## Remaining unowned in my area
`mission-feature-sync.ts` (1, a planning-lane check) and
`notification-service.ts` (1, *"has progressed past"* — a different
semantic needing its own thinking, not a mechanical swap). Taking those
next unless claimed.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
41031dbe2c