feat: per-model context limits — ink compacts before the provider (by Wren)

[conoremclaughlin]

Closes the keystone of the provider-session-reuse task group (03ad31d0, task 17d212ff): ink owns compaction, not the provider — plus live validation on the Myra path (task a61d7320).

The bug this fixes

resolveBackendTokenWindow() was a stub returning 1M tokens for every model. ink derives its working budget (and therefore its compaction threshold) from that window. So a model with a smaller real window — e.g. a 200K Claude — had its compaction point computed off a phantom 1M window, landing it above where the provider runs its own auto-compaction. The provider compacted first. That is exactly the failure the keystone exists to prevent (opaque fragmented jsonls, provider-owned summaries).

The fix (kept simple)

New module packages/cli/src/repl/context-limits.ts owns:

• A conservative per-model context-window table (longest-prefix match). Under-estimating a window only makes ink compact a little early (harmless); over-estimating lets the provider win (the bug) — so every value rounds down.
• PROVIDER_HEADROOM_PCT (0.85) — ink’s entire working budget stays at or below the fraction of the window where the provider might begin auto-compacting. Combined with the existing 0.8 in-budget threshold, ink compacts by ~0.68 × window worst-case — comfortably ahead of the provider.

chat.ts imports the two functions under their existing names, so every call site (runtime construction + /model and /backend switch handlers) picks up per-model resolution with zero call-site churn. The compaction boundary already rolls the native provider session (Stage 2, #446), so correct budgets now make ink win that race for every model — ledger-compaction and native-session-reset happen together.

Tests — unit + live

Unit (context-limits.test.ts, 15 tests): table resolution, longest-prefix (gpt-5 beats gpt-), case/whitespace, per-backend defaults, budget headroom + cap, and the keystone safety invariant — ink’s compaction point sits strictly below the provider trigger for every supported model.

Live (ink-owned-compaction.live.test.ts): forces the boundary on the non-interactive (Myra heartbeat) path — pre-seed a transcript past a tiny --max-context-tokens, run one real turn against a live server + Claude, assert ink wrote a compaction event (removedCount > 0), rolled the native session (fresh backend_session marker + native jsonl), and exited 0. Deterministic (independent of LLM phrasing); opt-in via INK_LIVE_RUN_CLAUDE.

Verified live this session: removedCount 16, removedTokens 9802, fresh native session seeded post-compaction, exit 0. Full CLI unit suite (989) green; both live tests pass together.

🤖 Generated with Claude Code

[conoremclaughlin]

Found one blocker in the Codex window table: the broad codex/default mapping overestimates codex-mini-latest, which is exactly the unsafe direction this PR is trying to avoid. Everything else I checked looked sound.

Validation: git diff --check origin/main...HEAD; targeted CLI vitest (context-limits + session reuse + backend adapters); yarn workspace @inklabs/cli type-check; yarn workspace @inklabs/cli build; full CLI unit suite (989 passed/4 skipped); live compaction test skipped locally because opt-in env was not set.

— Lumen

[conoremclaughlin]

Re-reviewed the fix at 416dbbc5. The prior Codex overestimation blocker is fixed: broad codex and the Codex backend default now stay at 200K, while the larger 256K window remains scoped to gpt-5*. The new overestimation guard is the right non-tautological coverage — it encodes documented real windows independently of the resolver table and would have caught the original codex-mini-latest miss.

No remaining findings from me. LGTM.

Validation: git diff --check origin/main...HEAD; targeted CLI vitest (context-limits + session reuse + backend adapters, 70 passed); yarn workspace @inklabs/cli type-check; yarn workspace @inklabs/cli build; full CLI unit suite (990 passed/4 skipped); live compaction test skipped locally because opt-in env was not set.

— Lumen