Fix WebMCP tool references and document MiMo/Luna retest

[jo32]

Problem and change

Models sometimes copy the catalog digest into a WebMCP tool revision or miscopy frame/document identifiers, causing avoidable rejected calls and recovery loops. Discovery now returns session-local toolRef handles; the browser plugin resolves each handle to the exact discovered identity and schema before dispatch.

References expire on tab/document/frame/version/schema changes. Legacy identity calls remain strictly validated. Errors distinguish not_dispatched from an unknown execution outcome and provide recovery guidance; uncertain actions are never automatically replayed. All runtime changes are in the Cordis browser plugin; upstream Harness and Jev changes are excluded.

Benchmark evidence

Retested six error-heavy WebMCP ON cases with MiMo V2.6 Flash and GPT-5.6 Luna, three attempts per case/model. All 36 formal attempts passed. Old baseline is one attempt per case; these tasks previously passed after tool failures, so this measures efficiency, not an increase in task success rate.

Six-case suite, old → new mean	MiMo	Luna
WebMCP call error rate	75.0% → 27.5%	38.9% → 21.6%
Agent time	20.04 → 5.75 min	3.07 → 3.02 min
Agent steps	108 → 60.3	46 → 41.7
Estimated USD, actual cache usage	$0.05645 → $0.02633	$0.04017 → $0.03624

This is a selected historical comparison, not a full-suite or contemporaneous A/B claim. Auth-boundary exploration regresses in steps/cost; Luna is also slower on auth-boundary and course enrollment. The report preserves these counterexamples and documents nine excluded MiMo configuration-calibration attempts and three Luna scoring-auth failures replaced by a complete three-attempt rerun. All 48 recorded attempts are included in the usage ledger. Luna costs are API-equivalent estimates, not subscription charges.

The committed report and sanitized per-attempt data include usage, rates, exclusions, operation-evidence checks, source hashes and raw-report hashes. Raw transcripts and credentials are not published. Paid benchmarks ran on the recorded working-tree snapshot before this PR was assembled; the seven code/test/doc files were copied unchanged onto current main.

Validation

On this isolated PR branch:

• Browser plugin type check, build and 162 tests across 17 files passed.
• Real Electron test passed native tool execution, catalog-digest rejection, stale-reference rejection after navigation and generated-tool version replacement.
• Exported costs were recomputed from recorded token usage; report links and staged diff checks passed.

Before PR isolation, the source workspace also passed pnpm check, pnpm build, and the full test suite with NODE_OPTIONS=--no-webstorage (needed for Node 25's localStorage behavior). These full-workspace checks were not rerun on the isolated PR branch.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
deepdeck	Ready	Preview	Sep 23, 2026 9:46am UTC