Post-0.65 batch 12: runner send states, worker program tool list, cargo test failures, #1441 step 4, gh relay launch route

Merged
UA
1h ago
+4396-429

Batch 12 on main 0a54c403 (after batch 11).

  • Runner notices: a not_delivered send reply is final for its send_id; completions resend under a new persisted id (at most 5), reminders drop; pending completions are followed with backoff up to 24 h (send.status when declared). Pins cortexkit-role-llm-runner =0.8.53. Real-daemon stand-ins now answer delivered sends per the 0.8.53 contract.
  • Tool catalog: a fetch with no preset on a worker scope, or for a code-mode run owned by a worker, renders the worker tool list; explicit presets, the presets list and policy freezing are unchanged.
  • cargo test compression keeps each failing test's panic message (bounded), every failure name and the test result line.
  • #1441 step 4: pass 4 resolves once per resolution key in member order, answers stored as ids; published segments byte-identical (exactness kit on neovim, superset, deno at two spill budgets). VS Code cold build peak 8.39 to 4.09 GiB, wall -29%.
  • gh relay: a bot write is judged by the shell's launching route when that route is still live with the captured stamp, so a stray route for the same session (OpenCode 1's after a head moved) no longer refuses it; other cases unchanged.
  • Integration tests launch aft from nextest's runtime path, with a guard against the build-time path.

Opened under the operator account because bot writes from moved heads are refused until the relay fix is live.


Summary by cubic

Batch 12 delivers runner send-state recovery, key-order pass 4 for #1441, cargo test panic retention, a gh relay launch-route fix, and the branch-keyed search descriptor.

Runner and tooling

  • A not_delivered runner reply is now final for its send id; completions resend under a new persisted id (at most 5 generations), then drop; pending sends are followed with backoff up to 24 h, surviving restarts.
  • Pins cortexkit-role-llm-runner to 0.8.53 and answers delivered sends per its contract.
  • A catalog fetch with no preset on a worker scope, or a code-mode run owned by a worker, now gets the worker tool list instead of the head list, including when served from a frozen policy; explicit presets are unchanged.
  • cargo test compression keeps each failing test's panic message within a bounded budget plus every failure occurrence and result line, with the final status reserved first.

Pass 4 and relay

  • Pass 4 resolves each distinct resolution key once in member order and stores answers as ids, with published segments byte-identical; VS Code cold build peak drops from 8.39 to 4.09 GiB and wall time by 29%.
  • The exactness kit now derives signature edits from Rust and Python receivers, reading multi-line signatures and picking a callee with a real signature parameter.
  • gh relay admits a bot write when the shell's launching route is still live with its captured stamp, so a stray route for the same session no longer refuses it; all other admission paths are unchanged.
  • Integration tests launch the aft binary from nextest's runtime path instead of the build-time one, with a guard against the build-time path.
  • The search descriptor marks the batch non-ranking: no changed path is in the ranking fence.