fix(provider): honor the host's tool set so compaction can summarize

[justin-carper]

Problem

Compaction fails on Cursor models:

Tool call not allowed while generating summary: cursor_context-mode_ctx_search

A failed compaction hard-errors the turn, so an oversized session cannot be compacted at all — the user is stuck.

Root cause

Three verified links:

1. The guard is opencode's. From the opencode 1.18.11 binary (SessionProcessor):

case "tool-input-start":
  if (a.assistantMessage.summary) throw Error(`Tool call not allowed while generating summary: ${c.name}`);
case "tool-call":
  if (a.assistantMessage.summary) throw Error(`Tool call not allowed while generating summary: ${c.name}`);

2. opencode declares zero tools on that turn. SessionCompaction.process builds the assistant message with summary: true and calls process({ ..., tools: {}, system: [] }). The bundled ai-sdk then drops the tool fields entirely — an empty tool record short-circuits before activeTools filtering:

function K1(K){ return K != null && Object.keys(K).length > 0 }
async function Q1({tools:K, toolChoice:Q, activeTools:Y}){ if(!K1(K)) return { tools: void 0, toolChoice: void 0 }; ... }

So the provider receives options.tools === undefined.

3. The provider never looked. grep -rn "\.tools\b|toolChoice|activeTools" src/ returned nothing before this change. Every turn ran the Cursor agent with its full native toolset, and cursorEventsToStream mapped that activity to provider-executed parts in the default "blocks" mode. The cursor_-prefixed generic name matches the reported cursor_context-mode_ctx_search exactly.

Chain: opencode declares no tools → Cursor agent uses its own tools anyway → provider forwards them as tool-input-start/tool-call → opencode's summary guard throws → compaction fails.

Why not restrict tools on the Cursor run instead

AgentOptions, LocalAgentOptions, and LocalSendOptions in @cursor/sdk@1.0.24 have no tool allowlist, denylist, or tool-disable field. mode: "plan" restricts writes but not reads/search. The only tools?: string[] in the package is a message field, not a run control. Suppression has to happen at the mapping boundary this plugin owns.

Fix

Honor the host's declared tool set. If opencode offered no tools, it cannot accept tool parts — so route that turn through the already existing and already tested toolDisplay: "reasoning" path, which renders Cursor's activity as reasoning text.

Turn	options.tools	Effective toolDisplay
Normal chat / subagent	non-empty array	configured value (default blocks) — unchanged
Compaction / summary	undefined	reasoning
Title generation	undefined	reasoning
activeTools filters everything out	[]	reasoning

The empty-array row matters: ai-sdk's early return only covers a null tool set, so a non-empty tools narrowed by activeTools arrives as [].

No new config knob, no env var, no changes to MCP forwarding, session pooling, or the skills mirror.

Why it surfaced now

Not a code regression in 0.7.1-next.0, but that release is what switched the failing path on. opencode merges the plugin's config-channel entry over its models.dev entry; Cursor models are absent from models.dev, so the fallback applies:

limit: { context: C.limit?.context ?? _?.limit?.context ?? 0, ... }

and auto-compaction is gated on exactly that value:

if (e.model.limit.context === 0) return !1;

git show be7a36e^:src/model-discovery.ts | grep limit → no matches. Pre-#89 the config channel carried no limit, so context resolved to 0 and auto-compaction could never fire for a Cursor model. #89 published real windows, so it now does.

Manual /compact has no such gate and was affected all along.

Compaction with an agent-backed provider

Worth recording, since Cursor keeps its own server-side conversation state: opencode-side compaction is genuinely effective here, and the existing pool design already produces the right lifecycle.

• The summary turn has system: [], so classifyTurn returns side-call → fresh ephemeral agent, pool untouched.
• The next real turn has collapsed history, so isStrictPrefix fails → divergence → resumeAgentId is never set → a fresh agent is seeded with the compacted transcript and re-pooled.

So the bloated Cursor agent is retired and context genuinely shrinks. The only thing that was broken is the tool-part leak on the summary turn.

Tests

507 → 514. Every new test was mutation-checked, not merely observed green:

• Revert both call sites to this.config.toolDisplay → the 2 no-tools regression tests fail.
• Gut the reasoning fold (tool parts suppressed but nothing emitted) → only the new positive assertions fail (expected '' to contain 'context-mode_ctx_search'), confirming they test the fold rather than mere absence.
• Drop || tools.length === 0 → the empty-array unit test fails.

Coverage: 4 unit tests for effectiveToolDisplay; a no-tools doStream emits none of tool-input-start/tool-input-delta/tool-input-end/tool-call/tool-result and folds the activity into reasoning while the summary text still reaches the host; a with-tools doStream still emits tool-call (positive control against silently killing normal tool blocks); a no-tools doGenerate mirror.

Verification

npx tsc --noEmit clean · npx vitest run 514 passed (35 files) · npm run build succeeds.

Not included

The EXC_GUARD / guarded-fd process kill observed shortly after a compaction is not addressed here and is not yet attributed to this plugin. A userland double-close cannot produce that signature — probed directly on Bun 1.3.5 and Node, where double close on file, socket, and stdio fds all yield EBADF, never EXC_GUARD — so it originates in native code, and the crash frames are unsymbolized. Investigation and its defensive hardening are deliberately held back rather than shipped on a guess.