refactor: consolidate Provider's four generate paths (plan 017)

[DavertMik]

Plan 017 — Consolidate the four duplicated generate paths in Provider

src/ai/provider.ts had four near-identical generate paths (chat, generateWithTools, generateObject, processImage) each re-implementing config assembly, usage recording, the idle-timeout race, and context-length recovery. This extracts four private helpers:

• recordUsage(agentName, modelName, usage) — single Stats.recordTokens call site (was 4). Also adds the cached token field to processImage (was the only path missing it).
• recoverFromContextLength(error, messages, options, retry) — single context-length reduce+retry, retry callback stays method-specific.
• buildGenerateConfig(defaults, overrides, options) — shared mergeProviderOptions + finalizeConfig assembly.
• raceWithIdleTimeout(fn, timeoutMs) — single Promise.race against the idle abort.

Drift reconciliation

This plan was written at 9e22b0d; since then plan 005 (abort in-flight request on idle timeout, PR #73) and plan 007 (PR #75) reshaped these methods. Per the plan's Step 4 guidance, plan 005's per-attempt AbortController + combinedAbortSignal is folded inside raceWithIdleTimeout — one controller per retry attempt (the helper is called inside the withRetry callback), signature (signal) => generateX({ ...config, abortSignal: signal }).

Behavior-preserving

• Config spread order is preserved per site: defaults land before ...(this.config.config) (so user config can still override the maxOutputTokens: 16384 default), overrides (model/abortSignal/stopWhen) land after ...options.
• generateWithTools no longer strips stopWhen from options — stopWhen: stopConditions in overrides already wins over ...options, and stopConditions still folds in options.stopWhen via extraStop. Identical result.

Deliberately NOT changed (behavior changes wanting their own review)

• chat() still has no idle-timeout race.
• processImage() still skips context-length recovery and keeps its minimal config path (routing it through buildGenerateConfig would change agent-level provider-option behavior).

Verification

• bun test tests/unit → 765 pass / 0 fail (+1 new context-length-recovery-through-helper test).
• bun test tests/integration/ → 63 pass / 0 fail.
• bun run lint → clean.
• tsc errors: 639 → 636 total (provider.ts 10 → 6) — refactor removed 4 promptTokens/completionTokens TS2339s; no new errors.
• Greps: Stats.recordTokens ×1, tryReduceMessages ×2, abortAfterIdle ×2, Promise.race ×1.

🤖 Generated with Claude Code

[github-actions]

Explorbot Self-Regression

Commit a0f91fa · run

Scenario	Result	Attempts	Duration
basic (native)	PASS	1/3	7m
experience: control	OK — failed as expected	1/1	1m
experience: seeded	PASS	1/3	1m

Attempt details

• basic (native) attempt 1 — PASS: login evidence: PASS (post-login plan=true, post-login research=true); research: PASS (files=6, wellFormed=true, keywords=7/3); scenarios: PASS (tests=6/5, features=4/3); tests passed: PASS (5 passed, 0 failed (reporter: 5 passed, 0 failed))
• experience: control attempt 1 — PASS: control: OK — failed as expected (0 passed, 1 failed)
• experience: seeded attempt 1 — PASS: seeded: PASS (1 passed, 0 failed)

---

Session analysis — basic (native):

Session Analysis

The Issues list feature was explored. Core flows for creating, viewing, searching, and filtering issues all function correctly. ET-2 had a verification retry that succeeded, and ET-5 required alternative selectors but achieved the goal.

Coverage

• Pages: /issues, /issues/:id, /issues?new=1, /issues?q=..., /issues?status=..., /issues?label=...
• Features: Create issue, View issue detail, Search by keyword, Filter by status, Filter by label

What works

• Create issue — ET-1 Create new issue via toolbar button
• View issue detail — ET-2 Open issue detail from list
• Search by keyword — ET-3 Search for existing issue by keyword
• Filter by status — ET-4 Filter issues by status using combobox
• Filter by label — ET-5 Filter issues by Bug label