fix(providers): send OpenCode session attribution header

[sambitcreate]

Summary

Every generation on the OpenCode Zen Go provider (opencode-go) failed before inference with:

400 {"type":"MissingSessionID","message":"Error from provider (Console Go):
Request is missing x-opencode-session and cannot be routed efficiently."}

OpenCode's gateway requires a stable per-conversation x-opencode-session header on every inference request (docs; the operators filed the same ask with jcode, Kilo Code, and VS Code — "any stable UUID per conversation works", "all outbound inference HTTP requests"). A Hermes control matrix proved the session header alone flips the request 400 → 200.

Aiden's pinned @earendil-works/pi-ai cannot emit that header — no published version can (upstream fixed it only inside the pi coding-agent, pi#4847; the pi-ai package fix pi#9326 is unreleased). So Aiden owns the attribution.

Changes

• New main/services/opencode-session-attribution.ts — attaches x-opencode-session: <conversation id> to the resolved runtime model. Gate: provider opencode / opencode-go / opencode-zen (forward-compat) or an exact opencode.ai base URL. Case-insensitive overwrite, clone-only (never mutates the shared catalog model), and ids are validated as bounded tokens so renderer-adjacent inputs can never carry CR/LF into a header.
• Model-level seam — resolveModelRuntimeWith / resolveModelRuntime / resolveBotModelRuntime / resolveCompactionModelMetadata grew an optional conversationId. Every pi-ai transport merges model.headers into the outgoing request, so one attachment covers chat turns, subagent children, isolated inference workers (headers survive the utility-process protocol), and the pinned-Bot dispatch path.
• Threaded ids — chat turns (chat.id), Bot preparation, compaction (both engines), chat titles, advisor dispatch, RPiV-BtW side questions, vision analysis, bot avatars (requestId), and dictation cleanup (fresh UUID for conversation-less one-shots).

Testing

• 15 new tests in main/services/opencode-session-attribution.test.ts: gate matrix (all provider ids, host matches, foreign/malformed URLs), blank/invalid ids, case-variant replacement, header preservation, catalog non-mutation, pinned-Bot dispatch, real wire-level delivery (local HTTP server + actual pi-ai completions transport), and native/custom resolver paths.
• Registered in test, test:google-provider, test:coverage.
• Verified locally: type-check clean; 201/201 across all 16 affected suites; npm run test:bots 440 pass. The 3 google-provider.test.ts failures on this Mac are pre-existing and environment-specific (verified identical on a stashed clean tree; CI is unaffected).

Notes

• Reviewed by two independent review agents: correctness/security (wire merge order, no cross-conversation leakage, no duplicate headers) and coverage/regression (all inference paths enumerated; the source-text contract in bots.contract.test.ts was fixed in this branch).
• No renderer/iOS/Android contract changes; no models.dev involvement.
• Deliberately out of scope (tracked in .memory/PROJECT-HISTORY.md): preserving provider error status/type in production diagnostics, a live gateway probe (incl. /messages + /responses routes), and removing this shim once a published pi-ai ships pi#9326.

[pullfrog]

ℹ️ No critical issues — one wire-coverage suggestion inline, plus a nitpick.

Reviewed changes

• New attribution module — main/services/opencode-session-attribution.ts clones the resolved runtime model and merges x-opencode-session, gated on provider id (opencode / opencode-go / opencode-zen) or an exact opencode.ai base-URL host, with a bounded id grammar and case-insensitive replacement of any existing variant.
• Resolver seam — resolveModelRuntimeWith, resolveModelRuntime, resolveBotModelRuntime, and resolveCompactionModelMetadata accept an optional conversationId and apply attribution on both the Pi-native and custom-provider paths.
• Threaded ids — chat turns (chat.id), Bot preparation (chat.id), both compaction engines, chat titles, advisor dispatch, BTW side questions, vision analysis, bot avatars (requestId), and dictation cleanup (fresh UUID) all pass a stable conversation or per-request id.
• Tests — 15 new cases in main/services/opencode-session-attribution.test.ts covering the gate matrix, id validation, clone-only behavior, pinned-Bot dispatch, both resolver paths, and real wire delivery; registered in test, test:google-provider, and test:coverage.

ℹ️ Nitpicks

• .papercuts/troubleshooting.md adds a subagent-artifacts/*_output.md note that is unrelated to this change; consider dropping it from this PR.

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using DeepSeek Flash ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes

• Pinned the session header on every OpenCode Go transport — the wire-level test now loops over openai-completions, openai-responses, and anthropic-messages, asserting per transport that the attributed request carries x-opencode-session while the unrelated provider does not.
• Dropped the unrelated .papercuts note — the subagent-artifacts/*_output.md troubleshooting entry is no longer part of this PR.

The new coverage is genuine: each lazy transport wraps its own streamSimple implementation, so a transport that stops merging model.headers fails its own assertion. Verified at c1c2c6f — 15/15 in the suite, tsc --noEmit and eslint clean.

  ^{｜ View workflow run ｜ Using DeepSeek Flash ｜ 𝕏}

[pullfrog]

All feedback from review #5185375806 was already addressed at the branch HEAD (c1c2c6f) — no new changes were required. Verified the wire test now pins the session header on all three OpenCode Go transports (main/services/opencode-session-attribution.test.ts:190-225), the unrelated .papercuts note was dropped, and the thread is resolved. Ran npx tsx --test main/services/opencode-session-attribution.test.ts: 15/15 pass. Fresh re-read of the review found no new comments.

  ^{｜ View workflow run ｜ via Pullfrog ｜ Using DeepSeek Flash ｜ 𝕏}