Skip to content

Open-Issue Engineering Triage

Snapshot recorded 2026-07-10 after reviewing every open GitHub issue and relevant upstream dependency state. GitHub remains the source of truth for current status.

There are 8 open issues. The queue is small, but it mixes real work, parked strategy, and historical notes.

Highest-value issues

#88 — AsyncMonty migration

This is becoming active: upstream released 0.0.19-beta.2 on July 9. Toolplane is protected by <0.0.19, but the migration is no longer hypothetical.

Why it matters:

  • dependency upgrade eventually becomes mandatory;
  • subprocess-per-session improves crash isolation;
  • it may affect session performance and snapshot behavior;
  • it may change the answer to #109’s concurrency problem;
  • the pool architecture aligns with the parked multi-tenant design.

Recommendation: begin a compatibility branch against the beta and run the full empirical session suite, but retain the production pin until stable upstream release.

#109 — Parallel calls inside snippets

This is the clearest product-performance item and directly supported by benchmark evidence. I would prioritize it more highly than the issue currently suggests.

Important sequencing: test AsyncMonty first. If it still serializes external futures, build bounded call_many; if upstream now provides useful concurrency, avoid duplicating machinery.

#113 — Benchmark follow-ups

The interesting pieces are:

  1. Codex cross-client run;
  2. real CLI+MCP composition task;
  3. genuinely code-resistant adaptive fixture.

The optional extra crossover repetitions are low value.

Recommendation: split this issue. Cross-client and composite workload are independently actionable and have different acceptance criteria.

Security issue with a cheap valuable slice

#108 — Security patterns

This contains three projects of very different maturity:

  • Discovery receipts: cheap, valuable now, and directly complementary to benchmark transcripts and audit logs.
  • Credential transport audit: worthwhile bounded security review.
  • Validate/execute hash binding: premature until Toolplane introduces code approval.

I would split it and implement discovery receipts first. Ideally each run records hashes/versions of the search results, schemas, and manifest the agent observed.

Small user-facing bug worth fixing

#93 — mcp status unexpectedly opens a browser

This violates an explicit read-only/browser-safe promise and leaves an orphaned OAuth page pointing at a dead callback. Even though recent import logic rewrites many mcp-remote configurations to direct OAuth, existing configurations remain affected.

This looks like a contained, user-visible correctness fix. A PATH shim for open/xdg-open, plus detection of the browser-open marker, is likely more robust than $BROWSER.

Strategically interesting, correctly parked

#96 — Remote/multi-tenant Toolplane

This is an excellent design note, not a work item. Its most important content is the foreclosure rules:

  • keep sessions, stores, credentials, grants, and audit ownership caller-scoped conceptually;
  • do not enable process-global state over multi-client transports;
  • carry identity explicitly through the bridge, as with run_id;
  • do not expose HTTP as though it were an authenticated enterprise service.

I agree strongly with leaving it parked until there is a real design partner.

Issue-hygiene candidates

#21 — Pyodide/pandas CDN test failure

Pyodide is now documented as supported but feature-frozen. The current issue should probably be resolved minimally:

  • mark the CDN test as network-dependent;
  • exclude it from default make test;
  • provide an explicit integration-test command.

Vendoring pandas wheels would conflict with the feature-frozen posture. This is a test-hygiene fix, not a backend investment.

#58 — Learnings

The remaining important insight is:

Monty’s safety advantage and its dialect tax are the same object.

That remains strategically relevant. But the issue body is malformed historical prose and is not actionable. I would move any unique material into docs and close it as archived learning.

Missing GitHub issues

The most consequential findings from the benchmark review are documented in Benchmark and Optimization Opportunities, but are not yet actionable GitHub items. Consider opening issues for:

  1. Hybrid facade experiment — direct tools plus code-mode tools under deferred loading.
  2. Native ToolSearch double-discovery reduction — current toolplane runs still use 1–2 more model requests.
  3. Immutable/counterbalanced benchmark harness — source fingerprinting, randomized arm order, exact request counts.
  4. Payload size and API-granularity benchmark axes.
  5. Many-server startup measurement — register_mcp_config() currently initializes servers sequentially.
  6. Longitudinal session/snapshot performance.
  7. 0.4.1 release and benchmark-headline reconciliation.

Suggested order

  1. Spike #88 against Monty beta.
  2. Resolve concurrency direction, then implement #109.
  3. Run Codex and composite-workload slices from #113.
  4. Benchmark the hybrid facade and discovery topology.
  5. Add discovery receipts from #108.
  6. Fix #93 and clean up #21/#58.
  7. Leave #96 parked.

The strongest existing issues are #88 and #109; the strongest missing issue is the hybrid facade experiment.