fix(gsd): prevent auto-mode setModel calls from persisting to settings.json

[deseltrus]

TL;DR

What: Fix three layers of model/thinking-level persistence leaking through auto-mode.
Why: Transient model switches silently overwrite the user's saved defaults — every subsequent session starts with the wrong model and thinking level.
How: (1) Pass options through the Extension API setModel wrapper, (2) add persist: false to auto-mode setModel calls, (3) thread persist through thinking level updates.

What

Three bugs combine to silently corrupt settings.json during auto-mode:

1. Extension API wrapper drops persist option (loader.ts) — ROOT CAUSE

The ExtensionAPI.setModel wrapper in loader.ts accepts only model, silently dropping the options argument:

// Before — options silently dropped
setModel(model) { return runtime.setModel(model); }

// After — options passed through
setModel(model, options) { return runtime.setModel(model, options); }

This means every pi.setModel(model, { persist: false }) call from any extension becomes runtime.setModel(model) — persist is always true. This single bug renders all downstream persist guards ineffective.

Note: The ExtensionCommandContext wrapper (agent-session.ts) already passes options correctly — only the loader.ts wrapper was affected.

2. Auto-mode setModel calls missing persist flag (auto.ts)

Two pi.setModel() calls omit { persist: false }:

• stopAuto() model restore (cleanup path)
• Hook model dispatch (transient per-unit override)

All 7 other setModel() calls in the GSD extension already use persist: false.

3. Thinking level not guarded by persist option (agent-session.ts)

_applyModelChange() calls setThinkingLevel() unconditionally. Even when the model switch uses persist: false, the thinking level change is always written to settings.json. If a transient switch triggers thinking level clamping (e.g. xhigh → high because an intermediate model lacks the capability), the clamped level is persisted.

File	Change
packages/pi-coding-agent/src/core/extensions/loader.ts	Pass options through setModel wrapper
packages/pi-coding-agent/src/core/agent-session.ts	Extract _applyThinkingLevel() that respects persist option
src/resources/extensions/gsd/auto.ts	Add { persist: false } to both setModel() calls
src/resources/extensions/gsd/tests/model-isolation.test.ts	Structural regression test

Why

The combination produces this failure mode:

1. User sets Model A with thinking level xhigh as their default
2. Auto-mode dispatches a unit, triggering a transient model switch via selectAndApplyModel
3. The extension calls pi.setModel(modelB, { persist: false })
4. loader.ts drops { persist: false } → model B is persisted to settings.json
5. The model switch triggers setThinkingLevel → if model B lacks xhigh support, thinking level is clamped to high and persisted
6. Next session reads corrupted defaults from settings.json

Related: #650 (original model config bleed report — the auto.ts fix addresses remaining call sites, but the loader.ts bug was the actual reason previous persist guards were ineffective)

How

Commit 1: auto.ts — add { persist: false } to remaining calls

Mechanical, consistent with the 7 existing calls that already use it.

Commit 2: agent-session.ts — thread persist through thinking level

Extract _applyThinkingLevel(level, options?) as a private method. Public setThinkingLevel(level) API unchanged.

Commit 3: loader.ts — pass options through wrapper (ROOT CAUSE)

One-line fix that makes all downstream persist guards actually work.

Test Evidence

• Root cause verified via runtime instrumentation: File-based debug logging in _applyThinkingLevel confirmed that persist=undefined reached the function even when callers passed { persist: false } — proving the loader wrapper was dropping the argument.
• Stack trace confirmed call path: selectAndApplyModel → loader.js:472 setModel(model) → runtime.setModel(model) — options lost at the wrapper boundary.
• Regression test: model-isolation.test.ts structural test verifies all setModel calls in auto.ts use persist: false. 8/8 pass.
• User-initiated changes still persist: Model selector and Ctrl+P cycling paths do not use persist: false and correctly write to settings.json.

---

AI-assisted contribution — reviewed and tested by contributor.

• fix

[github-actions]

🔴 PR Risk Report — CRITICAL

Files changed	4
Systems affected	4
Overall risk	🔴 CRITICAL

Affected Systems

Risk	System
🔴 critical	Agent Core
🔴 critical	State Machine
🔴 critical	Auto Engine
🟠 high	Extensions

File Breakdown

Risk	File	Systems
🔴	packages/pi-coding-agent/src/core/agent-session.ts	Agent Core, State Machine
🟠	packages/pi-coding-agent/src/core/extensions/loader.ts	Extensions
🔴	src/resources/extensions/gsd/auto.ts	Auto Engine
⚪	src/resources/extensions/gsd/tests/model-isolation.test.ts	(unclassified)

⚠️ Critical risk — please verify: state persistence, auth token lifecycle, agent loop race conditions, RPC protocol compatibility.

[jeremymcs]

Adversarial Review — persist: false thinking-level leak

Verdict: needs-attention

Finding (high severity)

The persist: false path added for transient auto-mode model switches is not fully isolated for thinking level.

_applyThinkingLevel (agent-session.ts:1748-1751) still appends thinking_level_change entries to durable session history even when options.persist === false. On session resume, replay logic calls setThinkingLevel(...) which writes back to settings by default — meaning a transient auto-mode model/thinking switch can survive in session state and silently overwrite the user's saved default in settings.json.

This defeats the intended isolation fix.

Recommendation

1. Don't record thinking-level changes as durable thinking_level_change entries when persist: false, or mark them as transient and skip them during session replay.
2. Ensure resume applies replayed thinking with a non-persisting path (e.g., internal setter with { persist: false }) so replay cannot mutate settings.json.

Suggested test

Add an integration test that:

1. Triggers setModel(..., { persist: false })
2. Reloads/resumes the session
3. Asserts settings.json default thinking level is unchanged

---

Review generated via Codex adversarial review

[jeremymcs]

This PR has merge conflicts with the base branch. Please rebase or merge main to resolve before review can proceed.

🤖 Automated PR audit — 2026-04-04