feat(harness): phase 2 parts 3–4 — deferred tool loading, unattended turns, mention tokens, routine run records

[aivsomkar]

Phase 2, parts 3 and 4 of the harness programme (docs/plans/2026-09-15-phase-2.md). Stacked on the part 2 PR. Spec, with every item marked built or deferred and why: docs/superpowers/specs/2026-09-15-phase-2-tools-and-turns-design.md.

In plain terms

Deferred tool loading (part 3), behind tools.deferred. With it on, a bot sees six core tools plus search_tools and use_tool instead of every tool schema on every call. It finds the rest by what they do and calls them by name; nothing depends on an engine honouring tools/list_changed, so it works on every engine. Off by default.

Turns that run unattended (part 4). Routine runs are told nobody is watching: take the most reversible reading, say which, and stop with a short failure summary when a sign-in, file, tool or permission is missing (decision 13). Instructions and delegation briefs may carry @name [[omb:type:id]] references that survive renames and resolve to one line with the id at run time. A recurring routine that comes due while its previous run is still going now counts the skip on the routine instead of running silently over it; overlap: "queue" lets the next run wait; a routine carries its failure streak. A bot's engine preference travels in a package export as setup intent (§16c).

Deferred, and why (in the spec)

The tool ladder (#1009: its first commit conflicts in four files on this chain, two of them UI; it stays its own PR to rebase), remote HTTP MCP servers (stdio-only through the registry, the probe and every driver's mount; a per-driver change to verify on real servers), structured ask_user (no harness tool exists to extend yet; #1184's option card is the surface), cancel_requested / dead_letter, F4's native structured output, and the five-section description lint over every tool (the four new tools follow it).

Measured

Proxy suite (a deferred child lists a core of under ten, finds list_rooms from "which rooms exist", calls it through use_tool, rejects a guessed name); mentions.test.ts; routines.test.ts (tokens resolved at run time with the definition untouched; a skipped occurrence counted and persisted; a queued second run; the failure streak counting and clearing); package-export.test.ts. The scorecard is unchanged because every switch is off by default.

Every engine

All harness-side (the proxy, prompt text, routine records, the export): full on every engine.

Platforms

Platform	Impact	Verified
macOS	server	suites, typecheck, lint, i18n
Windows	server	CI
iOS / Android	none	—

Checks

pnpm typecheck, pnpm lint, pnpm i18n:check clean; full vitest suite on this tip: 599 files, 7,182 passed, 54 skipped, 0 failed.

Live check (Sep 16, standalone instance, real Claude)

Deferred tools: asked which rooms exist, Claude called search_tools, then use_tool → list_rooms, and answered correctly; with the flag on, input per call fell from 43,634 (main) to 33,968 on T2 turn 1 (−22%), T1–T4 all correct. Mention tokens: a routine's run read "Worker (bot …)" with the resolved references block and completed. Record: docs/bench/scorecard/2026-09-16-phase2-live.md.

🤖 Generated with Claude Code

[vercel]

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

Project	Deployment	Actions	Updated
openmausbot-docs	Ready	Preview	Sep 16, 2026 2:33am UTC

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: b5660a6b-cde3-45b8-a13f-2615346ec2fd

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[milind-soni]

Closing at the maintainer’s request: we are deferring the Phase 2/3 harness stack to keep the current release scope smaller and reduce regression risk. This is a scope decision, not a finding that every change here is defective. The branch is being preserved so focused changes can be revisited separately.