rebase: give module agents the curated knowledge base

[tzhouam]

Problem

The rebase module agents could reach past failures (search_debug_memory) and learned skills (search_skills), but not the curated knowledge base — the human-written pages under knowledge/repos/vllm-omni/ covering components, models, CI and git lore. So an agent rediscovered by shell what a page already stated.

Measured on the v0.30.0 campaign, 526 tool calls in one run:

tool	calls
run_shell	169
read_file	114
edit_file	76
grep	26
search_debug_memory	17
search_skills	10
doc_search	2
doc_read	0

doc_search was already reachable on the harness path via the MCP tool bridge, so the gap was instruction, not capability: neither module prompt template mentioned the knowledge base at all, while the DEBUG_MEMORY tokens actively push agents toward search_debug_memory.

Change

• doc_search / doc_read added to the adapter tool schemas, appended after the parent dispatcher's 20 so that order stays an intact prefix.
• RebaseBackends gains both handlers, fail-closed by default like the rest, implemented in _build_backends over the same KnowledgeDocs view the review flows use (repo-scoped through the manifest's knowledge.repo_subdir).
• Guidance rides the builder-supplied ADAPTIVE_GUIDANCE slot, not a template edit — so the shipped templates stay byte-identical to the parent's copies and test_templates_are_parent_verbatim still passes. (I edited the templates first; that test caught it and I backed it out.)

Guidance tells agents to search the knowledge base before shell exploration, cite the page path when it informed a decision, and that where a page and the code disagree the CODE is authoritative for a rebase — say so in the decision file rather than editing to match the page.

Reviewer attention: goldens

Three goldens are deliberately regenerated — module_prompt_model_config.txt, module_prompt_worker_runner.txt, request_shape_turn1.json. The render and the tool surface both change by design, so no mechanism preserves them.

The prompt goldens no longer record the PARENT builder's output, and I rewrote that test's docstring to say so rather than let it keep claiming a parity it no longer has. The parent agent had no knowledge-base access, so this divergence is the point — but if retiring that parity needs an explicit decision, this is the line to object to.

Verification

• Full copilot suite green.
• doc_search returns real pages for the vllm-omni slice: components/scheduler/architecture.md, models/qwen-omni/_index.md, debug/guides/....
• Live-path render (live=True) carries the guidance, 24,567 chars, no unsubstituted tokens, and preserves supplied adaptive rules.

Honest limitation

Wired correctly, but on deepseek-flash the agents still did not call it — 0 doc_search in the first 99 tool calls of a run with this change live. Every layer verified (prompt render, schema, handler), so this is instruction-following, not wiring. Making it effective likely needs either a stronger tier or an enforcement point in the plan gate; that is follow-up, not this PR.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.