feat(pstack): support Claude Haiku 5.5 lanes

[jwerby]

What changed

Register Claude Haiku 5.5 as a pstack model family with low through max effort. Generate its five native agent definitions and add the family to the catalog ownership map. Add registry coverage for claude-haiku-5-5 external lanes.

This lets Codex use pstack's external Claude runner with claude-haiku-5-5 and preserves exact model verification in the receipt.

Verification

• bun test runner/registry.test.ts passes 31 tests.
• bun run typecheck passes.
• Ownership checks pass 30 tests.
• The full bun run static command has two five-second harness timeouts and one cleanup error. Both timeout tests pass alone with a 15-second timeout.
• The full runner suite has 496 passes, 6 timing failures, and 2 dangling-process errors under the default five-second test budget.
• Exact candidate installation and live behavior checks remain pending.
• The exact-head live-gate status remains pending.

[mergify]

Merge Protections

🟠 1 of 2 protections blocking · waiting on 🤖 CI

	Protection	Waiting on
🟠	Queue ready pull requests	🤖 CI
🟢	🚦 Auto-queue	—

🟠 Queue ready pull requests

Waiting for

• check-success=Unfret
• check-success=live-gate
• check-success=verify

Waiting checks: Unfret, live-gate, verify.

Automatically queue non-draft pull requests targeting main after verify, Unfret, and live-gate pass on the head and every review thread is resolved.

• check-success=Unfret
• check-success=live-gate
• check-success=verify
• #review-threads-unresolved = 0
• -draft

---

Show 1 satisfied protection

🟢 🚦 Auto-queue

When all merge protections are satisfied and these conditions match, this pull request will be queued automatically.

• -files~=^\.github/
• base = main

[greptile-apps]

[Medium impact] No new code defect was established; merge still requires the pending exact-head live check.

Summary

Adds Haiku 5.5 with five effort levels, generated Claude agents, and catalog ownership.

• Claude Haiku 5.5 is available as an optional pstack model with five effort levels.

_{Reviews (2) · Last reviewed commit: "test(verify): exercise Haiku native agen..." · Reviewed by Greptile}

[andrewfree]

This investigation used GPT-6 Astra and Claude Opus 5.5 against head 5fc2b5af4c88c0a39c43e018f2603cfb37279d98 and base 9368993ea3d96b491833d53c0964b389e868162c.

The catalog and native-agent addition addresses a real missing capability. The external runner already accepted the explicit claude-haiku-5-5 ID at base, with the same receipt model matching. Registration adds family metadata and constrains effort to low through max. The description should distinguish those changes.

The current registry suite passes 32 tests. Serial runner testing passed 353/353 and the verifier suite passed 167/167. Of six full parallel candidate package runs, five passed 503/503 and one had a child-reaping assertion failure at runner/run.test.ts:1497. Base package testing passed 502/502 in six runs. The files runner/run.ts, runner/run.test.ts and the provider adapters are unchanged. This PR changes registry data and a registry test; the failing case resolves a Codex lane through that registry. No causal link was established, and we have not reproduced the failure on base or identified the original reported timeouts. These checks used macOS arm64 and Bun 1.4.3, with real provider CLIs removed from PATH.

Could you attach the raw output or failing test names for the originally reported static timeouts, cleanup error, and "full runner suite" result of 496 pass / 6 timing failures / 2 dangling-process errors, along with the tested commit, Bun version, OS and exact commands? The body still contains those older counts.

The current verifier also rejects this PR before installation with Require an open same-repository PR, because its head is in jwerby/open-pstack. Its Claude transport uses headless sessions, which our operator policy prohibits. No live gate has been completed or published. Fork support and an authorized Claude app transport are needed before we can complete that gate from this environment.

I cannot edit this PR description from this account. Here is the suggested replacement for the author or a maintainer to apply.

Suggested updated PR description

What changed

Register the exact model ID claude-haiku-5-5 as an optional pstack family. Add five generated Claude-native agents for low, medium, high, xhigh and max effort, list all five in the verifier ownership map, and add pstack-haiku-5-5-max to the Claude native-agents recipe.

This enables Haiku selection through the setup catalog and named native-agent dispatch from Claude Code. The Codex-parent external runner already accepted the explicit model ID as an unregistered lane. Registration adds family metadata and restricts its allowed efforts to low through max; explicit ultra and default effort values are now rejected. Receipt model matching is unchanged.

This registers an exact Haiku 5.5 pin. It does not add a rolling haiku alias.

Verification

Independent investigation on 2026-10-18 used candidate 5fc2b5af4c88c0a39c43e018f2603cfb37279d98 and base 9368993ea3d96b491833d53c0964b389e868162c, on macOS arm64 with Bun 1.4.3. These checks used isolated copies and kept real provider CLIs off PATH.

• Registry tests passed 32/32 at the candidate and 31/31 at base.
• Generated catalog and native-agent drift check passed.
• Ownership checks passed 30/30.
• Serial runner suites passed 353/353 at the candidate and 352/352 at base.
• Verifier suites passed 167/167 in each of three runs at both revisions.
• Package strict typechecks and static invariants passed.
• Full parallel package testing is not consistently clean. Five candidate runs passed 503/503; one reported 502 pass and one failure in reaps the child and exits after repeated cancellation with a long deadline, at runner/run.test.ts:1497. Six base runs passed 502/502. The files runner/run.ts, runner/run.test.ts and the provider adapters are unchanged. This PR changes registry data and a registry test; the failing case resolves a Codex lane through that registry. No causal link was established. A pre-existing cleanup race is a hypothesis, not a reproduced base failure.
• The full bun run static command was not rerun as one command. Its static checks and verifier tests were run separately; the separate verifier typecheck remains unrun.
• Exact-candidate installation and live behavior verification remain pending.
• The exact-head live-gate status remains pending.

The initial report of two five-second static timeouts, one cleanup error, and a "full runner suite" result of 496 pass / 6 timing failures / 2 dangling-process errors predates the current head. The original failing-test names and raw output have not been supplied, so those failures have not been identified or dismissed.

The additional local checks used Bun 1.4.3 rather than the pinned Bun 1.4.0 and a restricted PATH. They do not replace the required installed live gate.