refactor(providers): share AbortCoordinator across both provider query loops

[griffinwork40]

What

openai-compatible/query.ts hand-rolled the per-session abort slot that anthropic-direct already encapsulates in AbortCoordinator — including six copies of the compare-and-clear invariant:

if (this.abortController === controller) this.abortController = null;

This moves AbortCoordinator to providers/shared/ and wires openai-compatible through it.

Why

That invariant is safety-critical: a leaked stale controller makes compact()'s idle guard misread "turn in flight" as idle and run summarization mid-turn. Maintaining it by hand in two providers is exactly the drift risk shared/tool-loop-cap.ts exists to prevent — and anthropic-direct/query.ts already has zero raw this.abortController references because it solved this.

AbortCoordinator was already provider-neutral: zero imports, no SDK types. Only its docstring said "Anthropic." Copying it into openai-compatible/ would have created the duplication this PR removes, so it moves to shared/ instead.

	before	after
raw abortController refs in openai query.ts	9	0
duplicated compare-and-clear sites	6	0 (one clear() write path)
abort-related fields on the class	4	1 collaborator

Four fields (abortController, pendingAbortReason, closeResolve, closedPromise) collapse into one abort collaborator. Net −26 lines of logic in query.ts; total file length is ~flat because the freed space went into documenting the contract at each site.

⚠️ One behavioural delta — intentional, convergent

compactHistory's beginAbort now goes through AbortCoordinator.begin(), which drains a pending abort reason.

An interrupt() parked between turns therefore pre-aborts the auto-compaction firing at that boundary and consumes the reason, instead of letting compaction run and deferring the abort to the next turn.

This matches anthropic-direct/query/compact-handler.ts:148 and is arguably the more correct reading (ESC should stop compaction too), but it is a change. Narrow repro window: pending interrupt and auto-compact enabled and the threshold crossing at that exact boundary. Documented inline at the call site.

Preserving the old behaviour would have required adding a non-draining begin variant to the shared class for one caller — working against the anti-drift goal of the PR. Flagging it explicitly rather than burying it.

Tests

Adds shared/abort-coordinator.test.ts (14 cases) — the class previously had no dedicated test, covered only indirectly via concurrent-session-isolation.test.ts. Now that it's load-bearing for two providers that gap was worth closing:

• scope lifecycle (begin → busy → clear → idle)
• stale-scope compare-and-clear protection (the highest-risk invariant)
• pending-abort drain, including drain-exactly-once and park-when-already-aborted
• the close-promise race the provider main loop depends on

Verification

• pnpm lint clean (tsc --noEmit, strict)
• Full suite: 699 files, 13082 passed, 14 skipped
• audit:sdk:check, audit:env:check, scan:env:check all green
• pnpm build succeeds

Scope

Part of #360 (god-class decomposition); follows #407, which extracted client/model-params/retry from this same file.

Deliberately not in this PR:

• SessionState bag — moves no logic on its own; only earns its place as the precondition for extracting the turn engine.
• Turn engine (_runTurnInner 209 LOC + runIteration 119) — the largest remaining mass and where ~4/5 of recent commits land, but a big diff on the hottest-churn path deserves its own review.
• compactHistory extraction — looks like a seam, isn't: its logic already lives in compact.ts behind injected callbacks, so the remaining 58 lines are a thin adapter. Moving it would be file-hop, not decomposition.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
agent-afk-docs	Ready	Preview, Comment	Jul 25, 2026 12:32am

[griffinwork40]

CI note: the Test (windows-latest) red is pre-existing, not from this PR

All gating checks pass — Test (ubuntu-latest) (the required gate) is SUCCESS, as are macOS, Lint & Build, Docs, Publish Bundle Smoke, and PTY.

windows-latest is advisory by design (ci.yml:94 sets continue-on-error for every non-ubuntu matrix OS, with the comment at ci.yml:83-89 noting the Windows-compat grind is in progress). Its failure does not block merge.

I verified it is not caused by this change:

	main @ e71fbce (this PR's base)	this PR
windows job conclusion	failure	failure
failing test files	85	86
failing tests	537	537

Same failing-test count, and the failures are the same pre-existing Windows-compat set — memory-loader.test.ts, skill-bridge.test.ts, bash.test.ts, tool-injector.test.ts, launchd.test.ts, etc. The two providers/ files that appear (anthropic-direct/grant-manager.test.ts, openai-compatible/auth.test.ts) also fail on main's Windows job and are untouched here.

Root cause of the sampled memory-loader.test.ts failure, for whoever picks up the compat work: the suite sets process.env['HOME'] to a tmp dir, but getAfkHome() falls back to join(homedir(), '.afk') (src/paths.ts:56) and Node's os.homedir() reads USERPROFILE on Windows — so the fixture directory is never consulted.

The file this PR adds passes on Windows: ✓ src/agent/providers/shared/abort-coordinator.test.ts (14 tests).

Worth flagging for future triage: a run-level conclusion of success on main does not mean the Windows job passed — continue-on-error masks it at the run level. The job-level conclusion has to be read directly.