feat: on-demand context-window variants in the /model picker

README.md

@@ -85,9 +85,16 @@ To delegate from another provider instead, just ask: *"Ask Claude to review this
 
 ## Context window controls
 
-Claude Code serves different context windows depending on the exact model id it receives (e.g. bare `claude-fable-5` serves 200K, while `claude-fable-5[1m]` serves 1M). `omp-claude-bridge` turns that into a single switch.
+Claude Code serves different context windows depending on the exact model id it receives (e.g. bare `claude-fable-5` serves 200K, while `claude-fable-5[1m]` serves 1M). `omp-claude-bridge` exposes both as **separate entries in the `/model` picker**, so you choose the window on demand:
 
-Set it once in `~/.omp/agent/claude-bridge.json`:
+- `claude-bridge/claude-opus-4-8` → **Opus 4.8 (1M)**
+- `claude-bridge/claude-opus-4-8-200k` → **Opus 4.8 (200K)**
+
+Switching window is just picking the other entry — no config edit, no reload. Every model appears once per window it supports, the `(1M)` / `(200K)` label is always shown, and each entry reports its true window so OMP's status bar and auto-compaction threshold stay accurate.
+
+### Default window
+
+The **unsuffixed** id (e.g. `claude-opus-4-8`) maps to a default window; the other window gets a `-1m` / `-200k` suffixed id. `provider.contextWindow` in `~/.omp/agent/claude-bridge.json` picks that default — it no longer hides models, it only decides which window is unsuffixed:
 
 ```json
 {
@@ -97,49 +104,56 @@ Set it once in `~/.omp/agent/claude-bridge.json`:
 }
 ```
 
-| Mode | Behavior |
+| Mode | Default (unsuffixed) window |
 | ---- | -------- |
-| `"auto"` *(default)* | Per-model policy based on measured SDK behavior. Respects `plan` and `longContextExtraUsage`. |
-| `"1m"` | Force the 1M window. Only 1M-capable models are registered; models with no 1M runtime are hidden. |
-| `"200k"` | Force the 200K window. Only 200K-capable models are registered; models with no 200K runtime are hidden. |
+| `"auto"` *(default)* | Per-model measured default. Respects `plan` and `longContextExtraUsage`. |
+| `"1m"` | 1M where the model has a 1M runtime, else its only window. |
+| `"200k"` | 200K where the model has a 200K runtime, else its only window. |
 
-The registered window always matches what the bridge actually requests, so OMP's status bar and auto-compaction threshold stay accurate.
+Both windows stay in the picker regardless of this setting (wherever a runtime exists); it only changes which one is the plain, unsuffixed id. So `modelRoles` / `enabledModels` that reference `claude-bridge/claude-opus-4-8` keep working and follow the default.
 
-### Per-model context window
+### Windows offered per model
 
-| Model | `auto` | `1m` | `200k` |
-| ----- | :----: | :--: | :----: |
-| `claude-opus-4-8` | 1M | 1M | 200K |
-| `claude-opus-4-7` | 1M | 1M | — hidden |
-| `claude-opus-4-6` | 200K¹ | 1M | 200K |
-| `claude-fable-5` | 200K | 1M | 200K |
-| `claude-sonnet-5` | 1M | 1M | 200K |
-| `claude-sonnet-4-6` | 200K² | 1M | 200K |
-| `claude-haiku-4-5` | 200K | — hidden | 200K |
+| Model | 200K entry | 1M entry | `auto` default |
+| ----- | :--------: | :------: | :------------: |
+| `claude-opus-4-8` | ✓ | ✓ | 1M |
+| `claude-opus-4-7` | — | ✓ | 1M |
+| `claude-opus-4-6` | ✓ | ✓ | 200K¹ |
+| `claude-fable-5` | ✓ | ✓ | 200K |
+| `claude-sonnet-5` | ✓ | ✓ | 1M |
+| `claude-sonnet-4-6` | ✓ | ✓ | 200K² |
+| `claude-haiku-4-5` | ✓ | — | 200K |
 
-¹ Opus 4.6 serves 1M in `auto` when `plan: "max"` or `longContextExtraUsage: true`.
-² Sonnet 4.6 serves 1M in `auto` when `longContextExtraUsage: true`.
+¹ Opus 4.6's `auto` default is 1M when `plan: "max"` or `longContextExtraUsage: true`.
+² Sonnet 4.6's `auto` default is 1M when `longContextExtraUsage: true`.
+
+The suffixed alternate exists only for the window that isn't the default — e.g. under `auto` you get `claude-opus-4-8` (1M) + `claude-opus-4-8-200k`, and under `"200k"` you get `claude-opus-4-8` (200K) + `claude-opus-4-8-1m`.
+
+> Forcing 1M is a *request*: some models may still be **served** 200K by your subscription entitlement. Set `CLAUDE_BRIDGE_DEBUG=1` to log the served window (see [Debugging](#debugging)).
 
 > An invalid `contextWindow` value logs a warning and falls back to `"auto"`, so a typo never breaks startup.
 
 ## Models
 
-Select any of these from `/model`:
+Pick any of these from `/model` — each entry shows a `(1M)` or `(200K)` label. The exact ids below assume the default `contextWindow: "auto"`; which id is unsuffixed vs `-1m` / `-200k` follows your configured [default window](#default-window).
 
-| Picker id | Notes |
-| --------- | ----- |
-| `claude-bridge/claude-fable-5` | Fast, 200K by default |
-| `claude-bridge/claude-opus-4-8` | Flagship, 1M by default |
-| `claude-bridge/claude-opus-4-7` | 1M by default |
+| Picker id (auto) | Window |
+| --------- | ------ |
+| `claude-bridge/claude-fable-5` | 200K |
+| `claude-bridge/claude-fable-5-1m` | 1M |
+| `claude-bridge/claude-opus-4-8` | 1M |
+| `claude-bridge/claude-opus-4-8-200k` | 200K |
+| `claude-bridge/claude-opus-4-7` | 1M |
 | `claude-bridge/claude-opus-4-6` | 200K (1M on Max / Extra Usage) |
-| `claude-bridge/claude-sonnet-5` | 1M by default, supports `xhigh` |
-| `claude-bridge/claude-sonnet-4-6` | 200K (1M on Extra Usage), supports `xhigh` |
-| `claude-bridge/claude-haiku-4-5` | 200K, cheapest |
+| `claude-bridge/claude-opus-4-6-1m` | 1M |
+| `claude-bridge/claude-sonnet-5` | 1M (supports `xhigh`) |
+| `claude-bridge/claude-sonnet-5-200k` | 200K |
+| `claude-bridge/claude-sonnet-4-6` | 200K (supports `xhigh`) |
+| `claude-bridge/claude-sonnet-4-6-1m` | 1M |
+| `claude-bridge/claude-haiku-4-5` | 200K (cheapest) |
 
 Bash commands issued by Claude Code get a 120-second default timeout (matching Claude Code's default), since OMP's bash has no timeout by default.
 
-**1M context notes:** Opus 4.7 and Opus 4.8 get 1M in `auto` by default. Opus 4.6 needs `plan: "max"` or `longContextExtraUsage: true`; Sonnet 4.6 needs `longContextExtraUsage: true`. Or just set `contextWindow: "1m"` to force it everywhere.
-
 ## AskClaude tool
 
 Available whenever the active provider is **not** claude-bridge. Your current model can hand work to Claude Code and wait for the result:

documentation/completed/features/2026-07-06_on-demand-context-window-variants.md

@@ -0,0 +1,87 @@
+# Task: On-demand context-window model variants (Option A)
+**Date Started**: 2026-07-06
+**Status**: Complete
+**Agent Lead**: 🏗️/⚙️
+
+## Plan
+
+### Objective
+Let the user choose a Claude model's context window **on demand** from OMP's native `/model` picker, by registering each model at every window it supports (1M and/or 200K) as a distinct, clearly-labeled entry. Switching window = pick the other entry. Keep `provider.contextWindow` from config as the **default** (it decides which window the plain, unsuffixed model id maps to). No new UI, no runtime command.
+
+### Current behavior (why there's nothing to select today)
+- Context window is a single **global** choice in `~/.omp/agent/claude-bridge.json` (`provider.contextWindow: auto|1m|200k`), read **once at load**; changing it needs a `/reload`. Not per-model, not on-demand.
+- OMP models carry a fixed `contextWindow`; there is no native per-model window knob (unlike thinking level's `setThinkingLevel`). Varying the window therefore requires registering variant models.
+- `src/models.ts` today: `applyLongContext` filters to **one** variant per model from the global mode; `claudeCodeModelId(model, settings)` maps a pi model id → Claude Code cli id using the global settings; `resolveForcedOneMRuntimeModel` / `resolveForcedTwoHundredKRuntimeModel` already hold each model's 1M/200K cli id (or `null` when unavailable).
+
+### Approach — variant scheme (per model, in MODEL_IDS_IN_ORDER)
+- **Available windows** = union of { 1M if `resolveForcedOneMRuntimeModel` ≠ null, 200K if `resolveForcedTwoHundredKRuntimeModel` ≠ null }. Independent of config, so both are pickable wherever a runtime exists.
+- **Default window** = the window `resolveClaudeCodeRuntimeModel(id, settings)` returns for the current config (auto respects `plan`/`longContextExtraUsage`; `1m`/`200k` force), falling back to the model's only available window when the preferred one has no runtime.
+- **Unsuffixed id** (e.g. `claude-opus-4-8`) = the default-window variant → preserves existing ids, roles, and the user's `config.yml`.
+- **Each other available window** gets a suffixed id: `${baseId}-1m` / `${baseId}-200k`.
+- **Display name** = `${baseName} (1M)` / `${baseName} (200K)` on every entry, for unambiguous labeling.
+- **`contextWindow`** on each entry = its true window (keeps status bar + auto-compaction accurate).
+
+Resulting picker (config = `auto`, Pro plan):
+
+| Label | Picker id | CC cli id | Window |
+| --- | --- | --- | --- |
+| Opus 4.8 (1M) | `claude-opus-4-8` | `claude-opus-4-8[1m]` | 1M (default) |
+| Opus 4.8 (200K) | `claude-opus-4-8-200k` | `claude-opus-4-8` | 200K |
+| Opus 4.7 (1M) | `claude-opus-4-7` | `claude-opus-4-7` | 1M (only) |
+| Opus 4.6 (200K) | `claude-opus-4-6` | `claude-opus-4-6` | 200K (default) |
+| Opus 4.6 (1M) | `claude-opus-4-6-1m` | `claude-opus-4-6[1m]` | 1M |
+| Fable 5 (200K) | `claude-fable-5` | `claude-fable-5` | 200K (default) |
+| Fable 5 (1M) | `claude-fable-5-1m` | `claude-fable-5[1m]` | 1M |
+| Sonnet 5 (1M) | `claude-sonnet-5` | `claude-sonnet-5[1m]` | 1M (default) |
+| Sonnet 5 (200K) | `claude-sonnet-5-200k` | `claude-sonnet-5` | 200K |
+| Sonnet 4.6 (200K) | `claude-sonnet-4-6` | `claude-sonnet-4-6` | 200K (default) |
+| Sonnet 4.6 (1M) | `claude-sonnet-4-6-1m` | `claude-sonnet-4-6[1m]` | 1M |
+| Haiku 4.5 (200K) | `claude-haiku-4-5` | `claude-haiku-4-5` | 200K (only) |
+
+Unsuffixed ids all keep their current default-per-config behavior, so `modelRoles.default: claude-bridge/claude-opus-4-8` still resolves and still gets 1M.
+
+### Code changes
+1. **`src/models.ts`**
+   - Add `parseVariantId(id)` → `{ baseId, forced?: "1m" | "200k" }` (suffix-aware; base ids never collide with `-1m`/`-200k`).
+   - Update `claudeCodeModelId(model, settings)`: if the id is suffixed, return the forced runtime's cli id (ignore global mode); else keep the existing default-per-config path.
+   - Add `buildVariantModels(models, settings)` that expands each model into its window variants (default-first ordering, labels, `contextWindow`). Replace `applyLongContext`; remove it (no back-compat shim).
+2. **`src/index.ts`**
+   - Line ~1634: register `buildVariantModels(MODELS, longContextSettings)` instead of `applyLongContext(...)`.
+   - The three `claudeCodeModelId` call sites (≈366 compact, ≈1251 streaming, ≈1463 AskClaude) stay unchanged — variant parsing is internal.
+3. **`tests/unit-*.mjs`** — expansion (ids/windows/cli ids/labels), default selection per config (auto/1m/200k × plan/extraUsage), fallbacks (opus-4-7 under `200k`, haiku under `1m`), no duplicate windows, `claudeCodeModelId` round-trips for suffixed + unsuffixed ids.
+4. **`README.md`** — update "Context window controls" + "Models": variants in the picker, config sets the default, keep the per-model table; note the entitlement caveat.
+
+### Tasks
+- [ ] models.ts: `parseVariantId` + variant-aware `claudeCodeModelId`
+- [ ] models.ts: `buildVariantModels` (default-first, labels, contextWindow); remove `applyLongContext`
+- [ ] index.ts: swap registration to `buildVariantModels`
+- [ ] Unit tests: expansion + id mapping + config defaults + fallbacks
+- [ ] README updates
+- [ ] `bun run typecheck` + `bun run test` green; manual `/model` check in OMP after `/reload`
+
+### Risks / notes
+- The default window has no suffixed alias (it's the unsuffixed id); a role hardcoded to that suffix breaks only when config flips that window to default. Common case (`auto`) is stable. Documented.
+- Entitlement: forced-1M variants may be **served** at 200K by the subscription (issue #18); `logServedContextWindow` already surfaces the gap. The picker offers the request; runtime may differ.
+- Picker grows 7 → ~12 entries; grouped default-first, acceptable.
+- `resolveModel` partial-match (AskClaude short names) unchanged: `"opus"` → first opus = default variant. Bonus: AskClaude can target `-200k` / `-1m` explicitly.
+- Single `registerProvider` retained → subagent re-registration guard unaffected.
+- User `config.yml` (`modelRoles.default`, `enabledModels: claude-bridge/*`) keeps working: unsuffixed id preserved, wildcard enables all variants.
+
+### Out of scope (MVP)
+- `/ctx` runtime command (Option B).
+- Removing the config knob (kept as default, per decision).
+- Per-window cost/thinking differences (none exist today).
+
+## Implementation Log
+
+### 2026-07-06 — Implemented (Option A), all checks green
+- **`src/models.ts`**: added exported `parseVariantId(id)` splitting `-1m`/`-200k` suffixes; made `claudeCodeModelId` variant-aware (suffixed id → forced runtime, ignores config; unsuffixed → config default). Replaced `applyLongContext` with `buildVariantModels` + a `variantName` helper — labels each entry `(1M)`/`(200K)`, sets the true `contextWindow`, orders default-first, offers both windows wherever a runtime exists, and makes the unsuffixed id the config default (falling back to the sole available window).
+- **`src/index.ts`**: import + registration now call `buildVariantModels(MODELS, longContextSettings)`. The three `claudeCodeModelId` call sites (compact ~366, streaming ~1251, AskClaude ~1463) unchanged — parsing is internal. `applyLongContext` fully removed (grep-confirmed no remaining refs).
+- **`tests/unit-context-window.mjs`**: rewritten for variants — 10 tests covering auto/Pro expansion (12 entries; ids/windows/labels), default-first ordering, auto/Max default flip (Opus 4.6), `200k` & `1m` config defaults + fallbacks (Opus 4.7, Haiku), the no-duplicate-window invariant, `parseVariantId`, and `claudeCodeModelId` round-trips (unsuffixed follows config, suffixed forces window, throws for impossible combos).
+- **`README.md`**: "Context window controls" + "Models" rewritten — variants in `/model`, `provider.contextWindow` now sets the *default* (no longer hides models), added "Windows offered per model" table + entitlement caveat, updated the picker-id table.
+- **Verification**: `bun run typecheck` clean; `bun run test` → 10/10 pass.
+- **Deployment**: the plugin is symlinked into OMP (`~/.omp/plugins/node_modules/omp-claude-bridge` → this repo), so the change goes live after a `/reload` in the OMP TUI.
+
+### Notes / follow-ups
+- Not done (out of scope, no explicit ask): version bump + CHANGELOG entry (still `0.7.0`). Adding variants changes the registered model set, so consider `0.8.0` before publishing.
+- Manual `/model` visual check must be done in the running OMP after `/reload` (can't drive the TUI from here).

src/index.ts

@@ -13,7 +13,7 @@ import { appendFileSync, mkdirSync, realpathSync, statSync } from "fs";
 import { homedir } from "os";
 import { dirname, join } from "path";
 import { PROVIDER_ID, messageContentToText, convertPiMessages } from "./convert.js";
-import { applyLongContext, buildModels, claudeCodeModelId, type ContextWindowMode, type LongContextSettings, resolveModel as _resolveModel } from "./models.js";
+import { buildVariantModels, buildModels, claudeCodeModelId, type ContextWindowMode, type LongContextSettings, resolveModel as _resolveModel } from "./models.js";
 import { MCP_SERVER_NAME, MCP_TOOL_PREFIX, extractSkillsBlock } from "./skills.js";
 import { verifyWrittenSession as _verifyWrittenSession } from "./session-verify.js";
 import { extractAllToolResults as _extractAllToolResults, type McpResult } from "./extract-tool-results.js";
@@ -1631,7 +1631,7 @@ export default function (pi: ExtensionAPI) {
 		longContextExtraUsage: providerSettings.longContextExtraUsage ?? false,
 		contextWindow,
 	};
-	const registeredModels = applyLongContext(MODELS, longContextSettings);
+	const registeredModels = buildVariantModels(MODELS, longContextSettings);
 
 	// Reset shared session on pi session lifecycle events
 	const clearSession = (event: string) => {