Best possible solution:

Land the shared, bounded native-tool index after refreshing it onto current main and confirming exact-head checks, while retaining deferred descriptions, full schemas, MCP entries, and overflow lookup inside the guest runtime.

Do we have a high-confidence way to reproduce the issue?

Not applicable as a conventional bug reproduction; the performance opportunity is demonstrated through an exact-branch live multi-provider benchmark against the merged predecessor.

Is this the best way to solve the issue?

Yes. Reusing the compact catalog formatter and exposing only a deterministic bounded index is narrower and safer than adding configuration or moving full schemas and untrusted descriptions into the model prompt.

AGENTS.md: found and applied where relevant.

Codex review notes: model internal, reasoning high; reviewed against a77c0770ffe6.

Label changes:

• add P2: This is a well-supported core agent performance improvement with limited blast radius and measurable workflow benefit.
• add proof: sufficient: Contributor real behavior proof is sufficient. The PR body contains after-fix live output from a real adversarial multi-provider harness showing improved success and substantial reductions in turns, calls, tokens, and latency with zero decoy calls.
• add rating: 🦞 diamond lobster: Overall readiness is 🦞 diamond lobster; proof is 🦞 diamond lobster and patch quality is 🦞 diamond lobster.
• add status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR. Sufficient (live_output): The PR body contains after-fix live output from a real adversarial multi-provider harness showing improved success and substantial reductions in turns, calls, tokens, and latency with zero decoy calls.

Label justifications:

• P2: This is a well-supported core agent performance improvement with limited blast radius and measurable workflow benefit.
• rating: 🦞 diamond lobster: Overall readiness is 🦞 diamond lobster; proof is 🦞 diamond lobster and patch quality is 🦞 diamond lobster.
• status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR. Sufficient (live_output): The PR body contains after-fix live output from a real adversarial multi-provider harness showing improved success and substantial reductions in turns, calls, tokens, and latency with zero decoy calls.
• proof: sufficient: Contributor real behavior proof is sufficient. The PR body contains after-fix live output from a real adversarial multi-provider harness showing improved success and substantial reductions in turns, calls, tokens, and latency with zero decoy calls.

PR surface:

Source +58, Tests +82, Docs +15. Total +155 across 4 files.

What I checked:

• Current-main gap: At current main, Code Mode keeps compact catalog metadata inside the guest runtime; the PR inserts the new prompt-index formatter after this point, so its central optimization is not already implemented. (src/agents/code-mode.ts:1269, a77c0770ffe6)
• Bounded prompt implementation: The branch constructs a deterministic OpenClaw/plugin-only index, marks result schemas unknown, omits descriptions, and limits the added index to 8,000 characters with an overflow fallback. (src/agents/code-mode.ts:1271, 271457894acc)
• Shared representation: The branch exports and reuses the existing compact catalog-entry formatter rather than creating a second input-hint implementation, keeping guest lookup and prompt metadata aligned. (src/agents/tool-search.ts:1261, 271457894acc)
• Regression coverage: Added tests cover exact IDs, deterministic ordering, deferred descriptions, unknown-output guidance, MCP exclusion, and bounded overflow behavior. (src/agents/code-mode.test.ts:422, 271457894acc)
• Documented contract: The documentation distinguishes hidden standalone tools from the bounded native-tool index and states that full schemas, descriptions, MCP entries, and overflow entries remain guest-side. Public docs: docs/reference/code-mode.md. (docs/reference/code-mode.md:39, 271457894acc)
• Live behavior proof: The PR reports an exact-branch, 72-tool adversarial matrix across OpenAI, Anthropic, and Google: validated tasks rose from 9/12 to 11/12, model turns fell from 49 to 29, and no decoy tool was called. (271457894acc)

Likely related people:

• steipete: Authored the merged compact-catalog optimization and this follow-up, giving the strongest recent history on the exact Code Mode performance path. (role: recent feature owner; confidence: high; commits: 57cc2ca308c1, 271457894acc; files: src/agents/code-mode.ts, src/agents/tool-search.ts, src/agents/code-mode.test.ts)
• Kaspre: Recent merged Code Mode work credited this contributor for model-visible exec-schema constraints and adjacent runtime error and hook behavior, making them a useful secondary reviewer for prompt-contract changes. (role: earlier Code Mode contributor; confidence: medium; files: src/agents/code-mode.ts, src/agents/code-mode.test.ts)

• 🦀 challenger crab: rare, exceptional readiness with strong proof, clean implementation, and convincing validation.
• 🦞 diamond lobster: very strong readiness with only minor maintainer review expected.
• 🐚 platinum hermit: good normal PR, likely mergeable with ordinary maintainer review.
• 🦐 gold shrimp: useful signal, but proof or patch confidence is still limited.
• 🦪 silver shellfish: thin signal; proof, validation, or implementation needs work.
• 🧂 unranked krab: not merge-ready because proof is missing/unusable or there are serious correctness or safety concerns.
• 🌊 off-meta tidepool: rating does not apply to this item.

Shiny media proof means a screenshot, video, or linked artifact directly shows the changed behavior. Runtime, network, CSP, and security claims still need visible diagnostics.

• ClawSweeper keeps one durable marker-backed review comment per issue or PR.
• Re-runs edit this comment so the latest verdict, findings, and automation markers stay together instead of adding duplicate bot comments.
• A fresh review can be triggered by eligible @clawsweeper re-review comments, exact-item GitHub events, scheduled/background review runs, or manual workflow dispatch.
• PR/issue authors and users with repository write access can comment @clawsweeper re-review or @clawsweeper re-run on an open PR or issue to request a fresh review only.
• Maintainers can also comment @clawsweeper review to request a fresh review only.
• Fresh-review commands do not start repair, autofix, rebase, CI repair, or automerge.
• Maintainer-only repair and merge flows require explicit commands such as @clawsweeper autofix, @clawsweeper automerge, @clawsweeper fix ci, or @clawsweeper address review.
• Maintainers can comment @clawsweeper explain to ask for more context, or @clawsweeper stop to stop active automation.