Model policy: six supervised Haiku cards for harvest, deep mapping and spec edits; record #1551

[AdaWorldAPI]

Haiku is removed from the forbidden-model list. It now handles cheap discovery and deep mapping of the substrate, which is too large to read on the main thread.

Cards (.claude/agents/, model haiku)

Card	Returns	Writes
harvest-scout	every candidate as path:line with the matching line	nothing
census-counter	raw counts, each with the exact command that produced it	nothing
quote-harvester	verbatim text with line numbers; FOUND / NOT FOUND checks	nothing
substrate-mapper	a fact map of one bounded unit (items, verbatim signatures and doc lines, names each body mentions, candidate callers)	its one map file
spec-editor	exact edits to one file, each applied only if its anchor occurs exactly once	the one target file
guarded-executor	a run-record for a pre-written command card	its run-record

Haiku still does not author content, draft, review, synthesize or decide.

Supervision

Defined in tiered-agent-execution-protocol.md § Haiku harvest tier.

• The orchestrator reads every item it relies on, re-runs one count per reply, spot-checks two items per map against the source, and reads the diff of every edit.
• A Haiku report of zero hits is never accepted as proof that something is absent.
• Hits and misses are logged in AGENT_LOG.md. A card that misses or concludes twice on one kind of task drops back to Sonnet for that kind.

First runs, verified by an independent grep

Search	Hits verified	Misses	Subagent tokens
lance-graph	9 of 9	1, a phrase wrapped across two lines, which the card had flagged as a limit	~444k
ndarray	2 of 2	0	~434k
tesseract-rs	7 of 7	0	~437k

Most of the token cost is the workspace context each subagent loads, not the search itself.

Also in this PR

• The D-HOLO: pre-register H11 and H12, add connectedness to H6b; record #1550 #1551 entries in the PR arc inventory and LATEST_STATE.
• Policy text updated in CLAUDE.md, both BOOT files, the agents README, worker-template-v2 and the cca2a skill.
• The review councils (5+3 and coresearch) keep their Sonnet floor, because their lenses synthesize.

Follow-up PRs in ndarray and tesseract-rs carry the same policy.

🤖 Generated with Claude Code

https://claude.ai/code/session_011BMJnDBAfzcuQh5HeT39Rh

---

Generated by Claude Code

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Warning

Billing warning: we have not been able to collect payment for this subscription for more than 72 hours. Please update the payment method or pay any pending invoices in Billing to avoid service interruption.

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.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-14T19:57:05.089585Z	f4e5d95	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.