fix: convert V2 goal plan tool input to JSON Schema

[abeisleem]

summary

• Convert PlanToolSchema to draft-7 JSON Schema for the V2 update_goal_plan tool instead of putting Zod instances directly into properties.
• Use input mode to preserve optional evidence and defaulted decisions; retain Zod parsing/refinements in the executor and leave V1 unchanged.
• Add a regression test for the serialized schema, required fields, nested arrays, nullable evidence, bounds, and strict objects. Update the tracked server bundle without unrelated compiler-renaming churn.

On OpenCode 2.0.25 with plugin 0.1.59, ordinary OpenAI Responses requests fail before generation:

invalid_function_parameters: Invalid schema for function 'update_goal_plan':
"optional" is not valid under any of the schemas listed in the 'anyOf' keyword

v2ObjectSchema(planToolArgs) only wraps the supplied properties; it does not convert their Zod types into JSON Schema. This affects ordinary prompts even when no goal is active because the registered tool catalog accompanies model requests.

testing

• Regression test fails on the old implementation and passes with the fix.
• Lint, typecheck, build, and pack dry-run pass.
• Focused V2 schema/plan tests: 3 pass, 0 fail.
• Isolated lifecycle smoke passes on OpenCode 2.0.25 using the built plugin and a deterministic local model; validated the update_goal_plan JSON Schema with Ajv in all 10 actual model requests. No live provider credentials or user goal state used. The temporary smoke harness adapts the existing script's older API routes/provider package name to 2.0.25; these unrelated harness changes are not in this PR. Not tested against a live OpenAI endpoint.
• Full suite: 360 pass, 2 fail (V1 paused-goal sampledAt clock mismatch and restart pending-attempt timeout). Untouched upstream also fails: 359 pass, 2 fail (restart pending-attempt timeout and V2 watchdog timing assertion). A separate V2 run showed watchdog timing failures too. These pre-existing/time-sensitive failures are outside this fix.

contribution disclosure

Implemented with OpenAI GPT-6.1 Sol through the OpenCode agent harness.

[danyel117]

The schema conversion is sound in the local review, and lint, typecheck, 362 tests, build, and pack checks pass. This PR overlaps #70 for the same provider-schema bug, so please coordinate on one final fix to avoid merging two implementations. If this PR is the chosen one, please add Closes #69 and the AI model/agent harness used (or state that the change was manual) to the description, as CONTRIBUTING.md requires. Its fork CI is also awaiting workflow approval.

[abeisleem]

Thanks for the review, @danyel117. I recommend consolidating into #70 so we can ship one fix promptly rather than duplicate the work. I have offered our nested-schema regression assertions and complementary OpenCode 2.0.25 wire-schema verification there: #70 (comment)

I have also added the required AI model/harness attribution to this description (OpenAI GPT-6.1 Sol through OpenCode). I have not added Closes #69 here because I am proposing #70 as the issue-closing PR.

Please confirm that #70 is the chosen implementation; I will then close this PR as superseded. Happy to help with any remaining verification.

[SimSonic]

DeepSeek the same:

[danyel117]

Thanks for the independent fix and schema regression coverage. PR #70 has now merged the provider-safe schema change and ported your nested-plan assertions with attribution, so I am closing this overlapping PR as superseded.