[factory] Dispatch attempt state is written under issueStateKey but read under issue.key, so maxAttempts never latches on GitHub

[khaliqgant]

Found while building a test for #435 (the lifecycle-terminal dispatch refusal). Filing separately rather than widening that PR: this changes dispatch retry semantics and belongs to whoever owns the attempt counters.

The mismatch

src/orchestrator/factory.ts:6039-6040, in the terminal-dispatch-failure branch:

await this.#recordDispatchFailure(decision.issue)
failedState = await this.#state.getDispatchAttempts(this.#workspaceId, decision.issue.key)

#recordDispatchFailure (factory.ts:9777) writes under issueStateKey(issue). The read one line later uses decision.issue.key. And issueStateKey (factory.ts:20242) is:

const issueStateKey = (issue: IssueRef): string =>
  githubIssuePathParts(issue.path) ? issueKey(issue) : issue.key

So the two agree on the Linear surface and disagree on the GitHub surface, where issueKey(issue) is a composite and issue.key is a bare number.

Consequence

On a GitHub-sourced Factory, failedState is always undefined, so:

const terminalFailure = liveStateChanged || Boolean(failedState?.terminal)

collapses to liveStateChanged alone. dispatch.maxAttempts therefore never drives the abandon-on-failure path for a GitHub issue — the only route to an abandoned row there is LiveDispatchStateChangedError.

Note the attempt row itself is still written correctly under issueStateKey, and #dispatchBlockReason (factory.ts:9745) reads it under the same key — so the retry-limit gate still works on a later sweep. What is lost is the in-pass decision: a dispatch that has exhausted its attempts is saved retryable instead of abandoned, and does not take the #recordDispatchTerminal / cleanup path at factory.ts:6092-6101 on that pass.

Reproduction

A GitHub-native issue, dispatch: { errorCooldownMs: 0, maxAttempts: 1 }, and a fleet client whose spawn fails. Expected abandoned; observed retryable:

AssertionError: expected [ 'retryable' ] to deeply equal [ 'abandoned' ]
-   "abandoned",
+   "retryable",

The identical test on the Linear surface produces abandoned, because there issueStateKey(issue) === issue.key.

Suggested direction

Read through issueStateKey(decision.issue) so the writer and reader agree — the same one-spelling rule #367 was filed for. Worth checking the other getDispatchAttempts call sites for the same shape while in there.

Not proposing a patch; flagging the disagreement and its blast radius.

[kjgbot]

Unowned, and it belongs to the identity theme

Triaged by the Factory lead. This has no owner, no labels, and no readiness label, so Factory cannot pick it up even when dispatch is working. Flagging rather than silently labelling it, because it changes retry semantics.

It is the same defect class as the design note in #228, Axis A: the same work unit is keyed differently depending on the surface it arrived by. Here the write uses issueStateKey(issue) and the read one line later uses issue.key; those agree on Linear and disagree on GitHub. That is the #211 shape in a new place, which is the argument for fixing the axis rather than the instances.

A lane brief is ready to run, so whoever picks this up does not have to re-derive it:

Deliverable. A PR fixing #438, with a test that fails before and passes after.

Definition of done.

1. A test reproducing the defect that fails on origin/main and passes with the fix. Quote both exit codes.
2. The equivalent Linear-surface test still passes — it already yields abandoned.
3. grep for other call sites pairing a write under issueStateKey with a read under issue.key or the reverse, and list them in the PR description. Do not fix them there; one work unit per PR.
4. Confirm CI with gh run list --branch <branch>, never --commit — the commit filter has returned empty for commits with green workflows, and an empty result is not a passing result.
5. Report every validation by exit code, never by absence of error.
6. Do not merge. mergePolicy is never.

A worktree is prepared at worktrees/factory-lane-438 on branch lane/438-attempt-state-key off origin/main.

I could not appoint a lane. agent-relay fleet spawn is not placing work from this session: untargeted spawns return status: dispatched with exit 0 and create no agent and no process, and targeted ones fail — --node chief-broker → exit 1 "target node is offline; placement TTL expired" (while the registry lists that node online with 26 live agents), and every other node name → exit 1 "Node not found". Recording it here because this surface is read.