[GaganSD]

• One search tool for web queries, public code, URL reading, and follow-up reading.
• Less context overhead: 16% lower estimated tool/prompt overhead by default, 44% lower with research enabled, and a 76% smaller representative result preview. These are offline character-based measurements, not measured billing savings.
• Local follow-up reads: revisit saved results and find passages without another network request.
• More search options: keyless URL reading through Jina Reader, optional Brave search, and recent-result reuse to reduce duplicate requests.
• Optional ranking and filtering: Jev is the default decision model; choose another model through /web-search classifier.
• 0.4.0 upgrade guidance is included in the changelog for custom tool references and settings.

[GaganSD]

Ponytail review of #90. Original review findings below.

Applied in b77424a after approval: removed all three identified legacy paths, migrated coverage without reducing the 432-test count, and removed 239 lines net (246 production-source lines). Release handoff remains intact. Local validation, 419 extracted-source regression tests, and installed Pi 1.4.0 loading passed.

Release preparation is included in 595cbf2: Unreleased changelog, breaking 0.3.1 → 0.4.0 migration guidance, and an exact post-merge preparation command. The existing checker rejects bundled feature/version changes and publication requires a merged release/next PR, so the actual release plan/version bump must remain a post-merge release-only PR. No workflow was dispatched and nothing was published.

Validation: 432 package tests, 25 tooling tests, lint, typecheck, production audit, clean-consumer packaging, 419 extracted-source regression tests, in-memory release generation, and offline loading against the operator's installed Pi 1.4.0 passed. All PR checks passed on 595cbf2.

• pi-web-search/src/index.ts:L56: delete: unified registration leaves web_search.ts, code_search.ts, and multi_search.ts with test-only callers; their three schemas and dispatch wrappers still ship in the tarball (161 lines). Port the applicable fixture coverage to search()/executeSearch() and delete the three obsolete modules; no compatibility wrappers are needed for unregistered tools.

• pi-web-search/src/search_read.ts:L247: delete: the zero-based searchRead() implementation and SearchReadSchema are used only by tests; registered ref calls already use searchReadRef(). Migrate those fixtures to one-based refs and remove the legacy reader/schema and its sole whole-result accessor, EvidenceStore.get() at src/evidence.ts:L83; keep the targeted listing/hit accessors and their safety coverage.

• pi-web-search/src/format.ts:L145: delete: removing the Coverage renderer leaves StreamResult.skipped without a reader. execute.ts still computes it through skippedSources() and threads it through judging/storage, but neither renderer details nor the structured schema exposes it. Remove that calculation/helper, its DEFAULT_CHAIN import, the unused type field, and the obsolete fixture field; retain the separate provider warnings and availability checks.

No ponytail: debt. Clean ledger.

net: -245 lines possible.