IronClaw
IronClaw tests deferred tool search against 1,000-tool catalogs
IronClaw proposes complete bounded signatures and namespace-aware previews for deferred tool discovery, backed by catalog baselines from 100 to 1,000 tools. The work treats discovery output as a model-facing contract, not a string dump that happens to fit today.
nearai/ironclaw PR #7410 is the inspected primary source: “feat(tool-search): return bounded complete signatures.” The related records below were inspected as supporting context rather than independent confirmation.
The facts
- nearai/ironclaw PR #7410 was created 2026-08-09T21:51:54Z and was open at inspection; its title is “feat(tool-search): return bounded complete signatures.” - nearai/ironclaw PR #7409 was created 2026-08-09T20:38:00Z and was open at inspection; its title is “test(tool-search): baseline catalogs at 100-1,000 tools.” - nearai/ironclaw issue #7405 was created 2026-08-09T19:59:36Z and was open at inspection; its title is “Improve deferred tool discovery with complete signatures and namespace-aware catalog previews.” - The repository reported 12596 stars, 1491 open issues, default branch main, and last push 2026-08-09T21:51:14Z. - The latest tagged-release baseline checked was ironclaw-v1.1.0 published 2026-08-06T08:39:11Z.
What changed
nearai/ironclaw PR #7410 reports: Summary Implements Phase 1 / Step 1 of #7405 as a PR stacked on #7409. Returns canonical parameters and schema_complete: true for ranked tool_search results whose schemas fit host-owned budgets. Preserves compact metadata with schema_complete: false when a schema exceeds either the 8 KiB per-result ceiling or the remaining 24 KiB response signature budget. Keeps BM25F ranking, authorization fitting, disclosure, and invocation behavior unchanged. Change Type [ ] Bug fix [x] New feature [ ] Refactor [ ] Documentation [ ] CI/Infrastructure [ ] Security [ ] Dependencies Linked Issue Related #7405 Stacked on #7409. Validation [x] cargo fmt --all -- --check [ ] Workspace-wide Clippy not run; affected crate passed cargo clippy -p ironclaw_loop_host --all-targets --all-features -- -D warnings [ ] cargo build — covered by tests and Clippy [x] Relevant tests pass: full cargo test -p ironclaw_loop_host (917 tests, 0 failures) and cargo test -p ironclaw_architecture_tests [ ] Integration feature: not applicable; no database-backed behavior changed [ ] Manual testing: not applicable; deterministic contract behavior is covered directly [ ] Agent review command
nearai/ironclaw PR #7409 reports: Summary Establishes PR 1 / Step 0 of #7405 without changing production tool-discovery behavior. Extends the existing 50-tool, 72-intent judged corpus with deterministic namespace-balanced distractors at 100, 500, and 1,000 tools. Commits retrieval-quality baselines at every scale and prints diagnostic index/query timings without turning host-dependent timings into flaky gates. Documents the end-to-end comparison arms, per-task result schema, required scenarios, rollout gates, and the boundary for any later orchestration work. Findings The current BM25F ranker remains stable as unjudged distractors scale the catalog from 100 to 1,000 tools. Recall@1 stays at 0.7865, MRR at 0.9492, and no-match accuracy at 1.0. Recall@10 moves from 0.9661 at 100 tools to 0.9557 at 500 and 1,000 tools; NDCG@10 moves from 0.9426 to 0.9404. This supports keeping BM25F as the retrieval authority while the following PRs improve the interaction around it. This is deliberately not an end-to-end model claim. The added tools are realistic deterministic distractors around the same 72 human-judged intents. Later slices must still run full-schema, current search/describe/call, complete-signature, namespace-summary, and pinned-profile arms while measuring task completion, discovery
nearai/ironclaw issue #7405 reports: Problem IronClaw's deferred-tool retrieval is strong, but the surrounding interaction still adds avoidable model turns and gives uneven catalog awareness at large tool counts. tool_search currently returns each match's name, capability ID, description, and required fields, while tool_describe returns the complete parameter schema. The model prompt therefore teaches a three-step tool_search -> tool_describe -> tool_call flow even when the top search result is unambiguous. The always-visible catalog preview is a byte-bounded alphabetical list of tool names. At large catalog sizes, early namespaces consume the preview and later integrations may not be represented. Catalog construction supports profile pins, but the production disclosure path currently supplies no pins. Do not replace IronClaw's bounded, authorization-fitted BM25F ranker with a simpler scorer. Improve the result contract, catalog presentation, pinning, and end-to-end evaluation around it. Status Landed on main: #7273: bounded schema-aware BM25F search over authorized tool metadata. #6958: progressive tool disclosure enabled by default. #7372: wide-catalog tool-schema token-reduction floor and drift reporting. Open implementation PRs: #7353: promotion-stable eager/deferred tool surfaces and Anthropic defer_loading / tool_reference encoding. This gives supported
Why it matters
Large tool catalogs fail twice: they waste context before selection, then omit the argument details needed to call what was found. Discovery must be both compact and complete enough for the next action. The operator test is whether visible output, retained state and authority still describe the same event after retries, background work or restart.
Current
The primary record was open when captured on 2026-08-10. The tagged-release baseline was ironclaw-v1.1.0 published 2026-08-06T08:39:11Z. Repository metadata, full source bodies, current pull requests and issues, releases, Google News, Hacker News, Lobsters, Metamesh and the rendered ClawCharts row were inspected. ClawCharts selected the subject; it did not prove the claim.
Evidence
The primary URL is https://github.com/nearai/ironclaw/pull/7410. Supporting records are https://github.com/nearai/ironclaw/pull/7409 and https://github.com/nearai/ironclaw/issues/7405. Source bodies, timestamps and states are preserved in the daily evidence bundle. Test counts and reproductions remain attributed to their authors unless identified as independently rerun.
Source boundary
Open work is described as open, closed work as closed, and operator reports as reports. A pull request is evidence of proposed or reviewed direction, not proof of a shipped release. Search residue, package mirrors and historically published source spines were excluded.
Operator take
Budget previews by namespace and signature, preserve required parameters and provenance, and benchmark retrieval quality as catalog size grows rather than counting only tokens saved. Preserve a before-state receipt, make the smallest reversible change, and verify the original failure independently.
Caveat
Public project records are mutable. Status, scope and evidence can change after publication. This brief records the inspected state and does not authorize changes to a reader’s deployment.
Source inspected; source state, environment and release boundary remain explicit.