feat: declare the Codex grammar through pi's constrainedSampling API

[zidou-kiyn]

Problem

createApplyPatchTool() attaches the Codex Lark grammar as a freeform property, but pi never reads that property. The provider request therefore always declares apply_patch as a plain JSON function tool, even for GPT models on providers that support OpenAI grammar tools.

Observed with pi 0.99.1, gpt-6.1-sol, and a Responses provider with compat.supportsOpenAIGrammarTools: true: the payload in before_provider_request contains { "type": "function", "name": "apply_patch" }.

Fix

Pi's public ToolDefinition.constrainedSampling hook accepts grammar variants (it is in the pinned 0.87.1 devDependency). This PR declares the existing grammar there:

constrainedSampling: { type: "grammar", variants: { openai_lark: APPLY_PATCH_LARK_GRAMMAR } },

• When the model's provider sets compat.supportsOpenAIGrammarTools (pi's bundled openai and openai-codex catalogs set it for GPT models), pi sends apply_patch as a native custom tool with format: { type: "grammar", syntax: "lark" }. pi maps the raw custom-tool input back to the single input parameter, so execute is unchanged.
• Everywhere else, pi falls back to the current function tool, so behavior there is unchanged.
• The freeform property is kept for existing consumers. The grammar, description, and schema are byte-for-byte unchanged.

Verification

• bun run check and bun run test pass (61 tests). The registration test now also asserts constrainedSampling.
• End to end with pi -p -ne -e ./src/index.ts on pi 0.99.1 + gpt-6.1-sol: all three requests declared custom:apply_patch (grammar), both patches (add file, then update line) applied, and every response was 200.

Note, not changed here

pi 0.99 prints on every start: Host-provided extension packages must be declared in peerDependencies with a "*" range, not dependencies: typebox. #40 moved typebox into dependencies on purpose for hosts that install with peer resolution disabled, so this PR leaves it alone. Flagging it in case you want to revisit that trade-off.

[code-yeongyu]

Thanks @zidou-kiyn, this is correct. I verified it against pi's own resolver at the pinned 0.87.1:

• With grammar support: resolveGrammarConstrainedSampling(createApplyPatchTool(), true) returns { format: "lark", inputProperty: "input" } with exactly APPLY_PATCH_LARK_GRAMMAR. The tool's schema meets pi's grammar requirement (one required string property).
• Without grammar support: it returns undefined, so providers without OpenAI grammar tools keep the plain function tool.
• On main: the first case fails, because the grammar is never declared.

I pushed one maintainer commit (79af773):

• It replaces the test assertion that pinned the constrainedSampling object with two behavioral tests through pi's public @earendil-works/pi-ai/api/constrained-sampling resolver, one per provider kind above.
• It adds a CHANGELOG entry crediting you.

I'll merge once CI is green on the new head.

[code-yeongyu]

Merged as f327003, with you credited as co-author and in the CHANGELOG. Thanks @zidou-kiyn: plain pi now sends apply_patch as a real grammar tool on providers that support it.