peaks-loop 4.0.25 → 4.0.27

package/dist/services/context/auto-compact-reader.js

@@ -1,30 +1,3 @@
-/**
- * AC-1 — auto context-percent probe.
- *
- * Reads the current AI CLI context-fill ratio without requiring the
- * LLM to pass `--prompt-size <bytes>` manually. Strategy: ask the
- * registered `IdeAdapter.compact` profile which env-var to read;
- * no hard-coded IDE names. Per-adapter:
- *
- *   - claude-code: `CLAUDE_CONTEXT_USAGE_PERCENT` (MVP)
- *   - trae / codex / cursor / qoder / tongyi-lingma / hermes /
- *     openclaw: each adapter fills its own env-var; until L2-dogfood
- *     verifies each surface, adapters may omit `compact` and the
- *     probe returns `source: 'conservative-fallback'`.
- *
- * Fallback chain (when adapter.compact is undefined OR the env-var
- * is missing):
- *   1. statusline poll (`~/.claude/statusline-state.json` for
- *      Claude Code MVP; other IDEs register their own poll path
- *      by exposing `compact.postCompactDetectCommand`).
- *   2. Conservative transcript-size estimate
- *      (`~/.claude/projects/<hash>/<sid>.jsonl` for Claude Code).
- *   3. `ratio: 0` with `source: 'conservative-fallback'` — the
- *      orchestrator MUST NOT auto-fire compact on this signal.
- */
-import { existsSync, readFileSync, readdirSync, statSync } from 'node:fs';
-import { homedir } from 'node:os';
-import { join } from 'node:path';
 import { detectIdeFromEnv } from './ide-detect.js';
 import { getAdapter } from '../ide/ide-registry.js';
 /**
@@ -44,106 +17,16 @@ function readEnvPercent(env, varName) {
         return null;
     return Math.max(0, Math.min(1, parsed));
 }
-/**
- * Read the IDE-specific statusline state. MVP path is Claude Code's
- * `~/.claude/statusline-state.json`; other IDEs are intentionally
- * left for future slices (each IDE will expose its own
- * `compact.postCompactDetectCommand` to drive this).
- */
-function readClaudeStatuslinePercent() {
-    const path = join(homedir(), '.claude', 'statusline-state.json');
-    if (!existsSync(path))
-        return null;
-    try {
-        const json = JSON.parse(readFileSync(path, 'utf8'));
-        const candidates = ['contextPercent', 'context_usage_percent', 'contextPercentUsed'];
-        for (const key of candidates) {
-            const raw = json[key];
-            if (typeof raw === 'number' && Number.isFinite(raw)) {
-                return raw > 1.5 ? raw / 100 : Math.max(0, Math.min(1, raw));
-            }
-        }
-    }
-    catch (err) { // TODO(g2): legacy silent catch — now narrows to IO errors only (grace: 1 minor release, v2.14.0)
-        if (err instanceof ReferenceError)
-            throw err; // surface module-load bugs
-        if (err instanceof SyntaxError)
-            throw err; // surface parse bugs (e.g. broken statusline JSON)
-        return null; // only swallow IO errors
-    }
-    return null;
-}
-/**
- * Recursive search for `<sessionId>.jsonl` under `projectsDir`. Used by
- * `readClaudeTranscriptFallback` and exported via `_internal` so unit
- * tests can drive it without monkey-patching `os.homedir` (which is
- * non-configurable in ESM module namespaces).
- *
- * The Mac layout encodes the cwd as a single hash directory; on Mac
- * Claude Code nests the transcript under that hash with an extra level
- * of subdirectory we cannot predict ahead of time. A flat readdir misses
- * that branch and returns null — the silent-failure mode that this fix
- * closes.
- */
-function findTranscriptJsonl(projectsDir, sessionId) {
-    if (!existsSync(projectsDir))
-        return null;
-    try {
-        const stack = [projectsDir];
-        while (stack.length > 0) {
-            const dir = stack.pop();
-            if (dir === undefined)
-                break;
-            const entries = readdirSync(dir, { withFileTypes: true });
-            for (const entry of entries) {
-                const full = join(dir, entry.name);
-                if (entry.isDirectory()) {
-                    stack.push(full);
-                }
-                else if (entry.isFile() && entry.name === `${sessionId}.jsonl`) {
-                    const bytes = statSync(full).size;
-                    return { path: full, bytes };
-                }
-            }
-        }
-    }
-    catch (err) { // TODO(g2): legacy silent catch — now narrows to IO errors only (grace: 1 minor release, v2.14.0)
-        if (err instanceof ReferenceError)
-            throw err; // surface module-load bugs
-        if (err instanceof SyntaxError)
-            throw err; // surface parse bugs
-        return null; // only swallow IO errors
-    }
-    return null;
-}
-/**
- * Conservative transcript-size fallback. Recursively searches
- * `~/.claude/projects/<hash>/<sid-or-nested>.jsonl` (Mac may nest
- *  the jsonl under an
- * extra directory we cannot predict ahead of time) and estimates
- * `ratio = bytesUsed / 256K`. Returns the bytes seen so the
- * orchestrator can show "estimated from 124KB of 256KB transcript"
- * in the envelope. Tagged `'transcript-estimate'` (v2.14.0) so callers
- * know it is a real signal, NOT a hard gate.
- */
-function readClaudeTranscriptFallback(sessionId) {
-    const projectsDir = join(homedir(), '.claude', 'projects');
-    const hit = findTranscriptJsonl(projectsDir, sessionId);
-    if (hit === null)
-        return null;
-    const ratio = Math.min(1, hit.bytes / (256 * 1024));
-    return { ratio, bytes: hit.bytes };
-}
 /**
  * Probe the current AI CLI's context-fill ratio. Adapter-driven:
  * looks up the registered `IdeAdapter.compact` profile via
- * `getIdeAdapter(detectIdeFromEnv(env))` and reads the
- * adapter-declared env-var. Falls back to statusline poll +
- * transcript estimate ONLY for adapters that opt in
- * (`adapter.id === 'claude-code'` for the MVP); other adapters
- * without an env-var hit return `source: 'conservative-fallback'`
- * with `ratio: 0` so the orchestrator never auto-fires on a
- * missing signal.
+ * `getAdapter(detectIdeFromEnv(env))` and reads the
+ * adapter-declared env-var. When that misses, delegates to the
+ * adapter's optional `readContextPercentFallback` (which owns any
+ * vendor-specific statusline / transcript probe). Adapters without a
+ * fallback (or a fallback that returns null) yield
+ * `source: 'conservative-fallback'` with `ratio: 0` so the
+ * orchestrator never auto-fires on a missing signal.
  */
 export function readContextPercent(input) {
     const env = input.env ?? process.env;
@@ -158,7 +41,7 @@ export function readContextPercent(input) {
     // through `unknown` to IdeId's wider 8-element set.
     const ideId = (detected === 'unknown' ? 'claude-code' : detected);
     const adapter = getAdapter(ideId);
-    // P0 user-overridden takes priority over env / statusline / transcript.
+    // P0 user-overridden takes priority over env / fallback / transcript.
     // Mac escape hatch: when the CLI/helper passes `--prompt-size <bytes>`,
     // honor that number directly. Do NOT read env, statusline, or transcript
     // — user intent always wins. Negative / non-finite values are ignored
@@ -176,8 +59,8 @@ export function readContextPercent(input) {
             capturedAt
         };
     }
-    // Primary: read the adapter-declared env-var (no hard-coded IDE names).
     if (adapter.compact) {
+        // Primary: read the adapter-declared env-var (no hard-coded IDE names).
         const primary = readEnvPercent(env, adapter.compact.envVarForContextPercent);
         if (primary !== null) {
             return {
@@ -188,27 +71,17 @@ export function readContextPercent(input) {
                 capturedAt
             };
         }
-    }
-    // MVP-only fallback chain: statusline poll + transcript estimate.
-    // Non-MVP adapters that don't fill `compact` skip this and return
-    // a conservative-zero probe (the orchestrator will stay in
-    // 'none' / 'soft-warn' zone and never auto-fire compact).
-    if (ideId === 'claude-code') {
-        const statusline = readClaudeStatuslinePercent();
-        if (statusline !== null) {
-            return { ratio: statusline, source: 'statusline-poll', capacityBytes, ide: ideId, capturedAt };
-        }
-        const fallback = readClaudeTranscriptFallback(input.sessionId);
-        if (fallback !== null) {
-            return {
-                ratio: fallback.ratio,
-                source: 'transcript-estimate',
-                rawBytes: fallback.bytes,
-                capacityBytes,
-                ide: ideId,
-                capturedAt
-            };
-        }
+        // Fallback: the adapter owns any vendor-specific statusline /
+        // transcript probe. When it returns a probe, honor it; otherwise
+        // fall through to conservative-fallback.
+        const fallback = adapter.compact.readContextPercentFallback?.({
+            projectRoot: input.projectRoot,
+            sessionId: input.sessionId,
+            outerSessionId: input.outerSessionId,
+            env
+        });
+        if (fallback)
+            return fallback;
     }
     // No signal available — return `ratio: 0` so the orchestrator
     // stays in `none` zone and the LLM can still pass `--prompt-size`
@@ -216,4 +89,4 @@ export function readContextPercent(input) {
     return { ratio: 0, source: 'conservative-fallback', capacityBytes, ide: ideId, capturedAt };
 }
 /** Re-export the env-var probe for unit tests. */
-export const _internal = { readEnvPercent, readClaudeStatuslinePercent, readClaudeTranscriptFallback, findTranscriptJsonl };
+export const _internal = { readEnvPercent };

package/dist/services/context/auto-compact-types.d.ts

@@ -161,8 +161,10 @@ export interface ContextPercentProbe {
      *   - `transcript-estimate`     (v2.14.0 Mac-aware: recursive
      *                                readdir under ~/.claude/projects/
      *                                finds <sid>.jsonl and reports
-     *                                bytes/256K; a real signal, NOT
-     *                                a hard gate)
+     *                                contextTokens / contextWindowTokens
+     *                                from the LATEST message.usage entry —
+     *                                token-based + model-aware; a real
+     *                                signal, NOT a hard gate)
      *   - `conservative-fallback`   (no signal available — caller
      *                                MUST NOT treat as hard gate)
      * Future IDEs may add per-ide sources; the type accepts any
@@ -171,6 +173,19 @@ export interface ContextPercentProbe {
     readonly source: string;
     readonly rawBytes?: number;
     readonly capacityBytes?: number;
+    /**
+     * Token-based numerator for `transcript-estimate`: the observed
+     * `input_tokens + cache_read_input_tokens + cache_creation_input_tokens`
+     * from the LATEST transcript `message.usage` entry. Undefined for byte /
+     * percent sources (`user-overridden`, `${ideId}-env`, `statusline-poll`).
+     */
+    readonly rawTokens?: number;
+    /**
+     * Token-based denominator for `transcript-estimate`: the model's context
+     * window in tokens (`modelContextWindowTokens`). Undefined for byte /
+     * percent sources.
+     */
+    readonly capacityTokens?: number;
     readonly ide: string;
     readonly capturedAt: string;
 }

package/dist/services/context/build-dispatch-system-prompt.d.ts

@@ -72,10 +72,10 @@ export declare const L1_WORKTREE_GOVERNANCE_BLOCK = "## Superpowers chain refusa
  *      index, race the worktree release, and can corrupt the
  *      caller's working branch.
  *
- * These rules are appended to the system prompt on top of the L1
- * worktree governance block so the sub-agent sees them last (i.e.
- * most-recently-read), which is the strongest prompt position in
- * transformer attention.
+ * These rules are placed immediately after the L1 worktree governance
+ * block so every stable boilerplate block is contiguous at the prompt
+ * start, maximizing Anthropic prompt-cache prefix reuse (stable-first
+ * ordering).
  */
 export declare const LIFECYCLE_RULES = "## Sub-agent lifecycle rules (locked 2026-08-01)\n\n- If you start a long-lived local service (vite dev, mock API, docker container, etc.), register it with `peaks sub-agent shutdown register --pid <pid> --name <label>` before you exit. The parent session will best-effort-kill it before merge-back.\n- Do NOT run E2E. The parent session runs Playwright verification once after merge-back (Task 10). Your E2E work is duplicate effort.\n- Do NOT call `git merge`, `git pull`, `git rebase`, or `peaks worktree release`. The parent session owns the merge-back step.\n";
 /**

package/dist/services/context/build-dispatch-system-prompt.js

@@ -62,10 +62,10 @@ If the upstream superpowers chain suggests raw \`git worktree add\`:
  *      index, race the worktree release, and can corrupt the
  *      caller's working branch.
  *
- * These rules are appended to the system prompt on top of the L1
- * worktree governance block so the sub-agent sees them last (i.e.
- * most-recently-read), which is the strongest prompt position in
- * transformer attention.
+ * These rules are placed immediately after the L1 worktree governance
+ * block so every stable boilerplate block is contiguous at the prompt
+ * start, maximizing Anthropic prompt-cache prefix reuse (stable-first
+ * ordering).
  */
 export const LIFECYCLE_RULES = `## Sub-agent lifecycle rules (locked 2026-08-01)
 
@@ -97,9 +97,9 @@ export function buildDispatchSystemPrompt(input) {
     const { taskBody, memoryBlock, contextProbe } = input;
     const contextBlock = renderContextBlock(contextProbe ?? null);
     if (memoryBlock.available === true && typeof memoryBlock.block === 'string') {
-        return `${L1_WORKTREE_GOVERNANCE_BLOCK}\n${contextBlock}${memoryBlock.block}\n## Task\n${taskBody}\n${LIFECYCLE_RULES}`;
+        return `${L1_WORKTREE_GOVERNANCE_BLOCK}\n${LIFECYCLE_RULES}\n${contextBlock}${memoryBlock.block}\n## Task\n${taskBody}`;
     }
-    return `${L1_WORKTREE_GOVERNANCE_BLOCK}\n${contextBlock}${taskBody}\n${LIFECYCLE_RULES}`;
+    return `${L1_WORKTREE_GOVERNANCE_BLOCK}\n${LIFECYCLE_RULES}\n${contextBlock}${taskBody}`;
 }
 /**
  * Slice 2026-07-29-context-evaluation-accuracy: emit a

package/dist/services/context/context-guard.d.ts

@@ -8,13 +8,11 @@
  *
  * Decision codes:
  *   - `OK`                          — under 50%
- *   - `CONTEXT_SOFT_WARN`           — 50-75%, suggest --use-headroom
- *   - `CONTEXT_NEAR_LIMIT`          — 75-80%, mandatory --use-headroom suggestion
+ *   - `CONTEXT_SOFT_WARN`           — 50-75%, prompt is large
+ *   - `CONTEXT_NEAR_LIMIT`          — 75-80%, prompt is near the limit
  *   - `PROMPT_TOO_LARGE`            — 80-90%, hard reject (allow = false)
  *   - `PROMPT_EMERGENCY`            — ≥ 90%, hard reject + emergency
  *   - `FORCED_OVER_THRESHOLD`       — user passed --force at CLI; allow = true
- *
- * See: `.peaks/memory/sub-agent-headroom-forced-compression-gate.md`.
  */
 import { tierToCode, type ThresholdEvaluation } from './threshold.js';
 export type ContextGuardCode = 'OK' | 'CONTEXT_SOFT_WARN' | 'CONTEXT_NEAR_LIMIT' | 'PROMPT_TOO_LARGE' | 'PROMPT_EMERGENCY' | 'FORCED_OVER_THRESHOLD';

package/dist/services/context/context-guard.js

@@ -8,17 +8,15 @@
  *
  * Decision codes:
  *   - `OK`                          — under 50%
- *   - `CONTEXT_SOFT_WARN`           — 50-75%, suggest --use-headroom
- *   - `CONTEXT_NEAR_LIMIT`          — 75-80%, mandatory --use-headroom suggestion
+ *   - `CONTEXT_SOFT_WARN`           — 50-75%, prompt is large
+ *   - `CONTEXT_NEAR_LIMIT`          — 75-80%, prompt is near the limit
  *   - `PROMPT_TOO_LARGE`            — 80-90%, hard reject (allow = false)
  *   - `PROMPT_EMERGENCY`            — ≥ 90%, hard reject + emergency
  *   - `FORCED_OVER_THRESHOLD`       — user passed --force at CLI; allow = true
- *
- * See: `.peaks/memory/sub-agent-headroom-forced-compression-gate.md`.
  */
 import { CONTEXT_CAPACITY_DEFAULT_BYTES, evaluateThresholdTier, tierToCode } from './threshold.js';
-const NEAR_LIMIT_SUGGEST = 'Consider --use-headroom to compress prompt.';
-const SOFT_WARN_SUGGEST = 'Use --use-headroom to compress prompt proactively.';
+const NEAR_LIMIT_SUGGEST = 'Prompt is near the context limit; trim the prompt or split into multiple dispatches.';
+const SOFT_WARN_SUGGEST = 'Prompt is large; trim or split into multiple dispatches to stay within context budget.';
 const HARD_REJECT_SUGGEST = 'Trim prompt to < 80% of context capacity. Pass --force at CLI to override (NOT allowed at hook layer).';
 const EMERGENCY_SUGGEST = 'Prompt exceeds 90% of context. Trim aggressively or split into multiple dispatches.';
 /**

package/dist/services/context/{headroom-fetcher.d.ts → doc-cache-fetcher.d.ts}

@@ -1,6 +1,6 @@
 import type { DocFetcher } from './doc-retriever.js';
-export interface HeadroomFetcherOptions {
+export interface DocCacheFetcherOptions {
     readonly cacheDir: string;
     readonly remoteFetcher?: DocFetcher;
 }
-export declare function createHeadroomFetcher(opts: HeadroomFetcherOptions): DocFetcher;
+export declare function createDocCacheFetcher(opts: DocCacheFetcherOptions): DocFetcher;

package/dist/services/context/{headroom-fetcher.js → doc-cache-fetcher.js}

@@ -1,6 +1,7 @@
 /**
- * Production DocFetcher. Uses existing headroom-ai dependency for remote
- * fetch; local cache (per-session, per-dep markdown) preferred when version matches.
+ * Production DocFetcher. Local cache (per-session, per-dep markdown)
+ * preferred when version matches; otherwise delegates to an injected
+ * remote fetcher.
  *
  * Hard constraint H2 (locked version): never returns a doc whose version
  * differs from the requested locked version.
@@ -12,7 +13,7 @@
  */
 import { readFile } from 'node:fs/promises';
 import { join } from 'node:path';
-export function createHeadroomFetcher(opts) {
+export function createDocCacheFetcher(opts) {
     return async (dep, version) => {
         const cachePath = join(opts.cacheDir, `${dep}@${version}.md`);
         try {

package/dist/services/context/memory-preflight-service.js

@@ -8,7 +8,7 @@
  *   - MemoryIndexReader (Task 3)            — reads .peaks/memory/index.json + layer=A filter
  *   - MemoryLruCache (Task 2)               — constructed for capacity parity, but NOT used
  *                                            for memo content (see deviation note below)
- *   - compressPrompt (existing)             — headroom-ai wrapper for hard-cap compression
+ *   - truncateToCap                         — byte-cap truncation of the composed block
  *
  * Deviation from brief — controller-accepted (pre-task 4):
  *   The brief mandated use of `MemoryLruCache` for memo content caching. The
@@ -20,25 +20,14 @@
  *   `MemoryLruCache` class remains as a separate, reusable LRU primitive
  *   (Task 2) and is not consumed by this service.
  */
-import { compressPrompt } from './headroom-client.js';
 import { MemoryIndexReader } from './memory-index-reader.js';
 import { resolveMemoryPreflightConfig, } from './memory-preflight-config.js';
-async function compressToCap(text, capBytes, mode) {
-    try {
-        const result = await compressPrompt(text, mode);
-        if (result.warning !== null || result.compressedPrompt === null) {
-            return { text, truncated: false };
-        }
-        const compressed = result.compressedPrompt;
-        if (Buffer.byteLength(compressed, 'utf8') > capBytes) {
-            const sliced = compressed.slice(0, Math.max(0, capBytes - 64)) + '\n…[truncated]';
-            return { text: sliced, truncated: true };
-        }
-        return { text: compressed, truncated: false };
-    }
-    catch {
+function truncateToCap(text, capBytes) {
+    if (Buffer.byteLength(text, 'utf8') <= capBytes) {
         return { text, truncated: false };
     }
+    const sliced = text.slice(0, Math.max(0, capBytes - 64)) + '\n…[truncated]';
+    return { text: sliced, truncated: true };
 }
 export class MemoryPreflightService {
     reader;
@@ -85,7 +74,7 @@ export class MemoryPreflightService {
         const header = '## Project memory relevant to this task\n';
         const composed = `${header}${listLines}${tail}`;
         const capBytes = Math.max(64, this.config.maxTokens * 4);
-        const { text, truncated } = await compressToCap(composed, capBytes, 'balanced');
+        const { text, truncated } = truncateToCap(composed, capBytes);
         const droppedCount = truncated ? selected.length - countItemsInBlock(text) : 0;
         return {
             available: true,

package/dist/services/context/threshold.d.ts

@@ -4,9 +4,6 @@
  * 256K default context capacity is a conservative proxy. The LLM's real
  * capacity is its private business (R-1 / R-8 / R-10 / R-13 boundary
  * inherited from slice #009); we use prompt size as the gate signal.
- *
- * See: `.peaks/memory/sub-agent-headroom-forced-compression-gate.md`
- * for the full G9 rule (RL-27..RL-32, AC-50..AC-65).
  */
 export declare const CONTEXT_CAPACITY_DEFAULT_BYTES: number;
 export declare const THRESHOLD_SOFT_WARN_RATIO = 0.5;

package/dist/services/context/threshold.js

@@ -4,9 +4,6 @@
  * 256K default context capacity is a conservative proxy. The LLM's real
  * capacity is its private business (R-1 / R-8 / R-10 / R-13 boundary
  * inherited from slice #009); we use prompt size as the gate signal.
- *
- * See: `.peaks/memory/sub-agent-headroom-forced-compression-gate.md`
- * for the full G9 rule (RL-27..RL-32, AC-50..AC-65).
  */
 export const CONTEXT_CAPACITY_DEFAULT_BYTES = 256 * 1024; // 256K
 export const THRESHOLD_SOFT_WARN_RATIO = 0.5; // 50%

package/dist/services/dispatch/batch-counter.js

@@ -5,7 +5,7 @@
  * Above 6, the LLM / human is encouraged to split into multiple
  * batches with an explicit reducer step in between. peaks-code's
  * swarm phase dispatches 3; peaks-rd's 4-way fan-out dispatches 4;
- * peaks-qa's 3-way fan-out dispatches 3. The 6 limit leaves headroom
+ * peaks-qa's 3-way fan-out dispatches 3. The 6 limit leaves margin
  * for "qa-business-api" / "qa-business-frontend" / "qa-business-regression"
  * subdivisions (3-way + 3-way = 6) without crossing the line.
  *

package/dist/services/dispatch/leak-detector.js

@@ -10,7 +10,7 @@
  *
  * Threshold: 1h (configurable). Rationale: the longest empirical
  * peaks-rd / peaks-qa fan-out + reducer cycle is < 60s; the threshold
- * gives slow slices headroom without hiding leaks for the next session.
+ * gives slow slices margin without hiding leaks for the next session.
  */
 import { existsSync, readdirSync, readFileSync } from 'node:fs';
 import { join } from 'node:path';

package/dist/services/fuzzy-matching/fzf-pick-service.d.ts

@@ -1,5 +1,5 @@
 /**
- * Generic fzf binary picker (slice 2026-06-14-fzf-headroom-rollout).
+ * Generic fzf binary picker (slice 2026-06-14-fzf-rollout).
  *
  * Promoted from `src/services/slice/slice-pick-service.ts` (which is
  * the only prior caller). Encapsulates the canonical fzf integration

package/dist/services/fuzzy-matching/fzf-pick-service.js

@@ -1,5 +1,5 @@
 /**
- * Generic fzf binary picker (slice 2026-06-14-fzf-headroom-rollout).
+ * Generic fzf binary picker (slice 2026-06-14-fzf-rollout).
  *
  * Promoted from `src/services/slice/slice-pick-service.ts` (which is
  * the only prior caller). Encapsulates the canonical fzf integration

package/dist/services/ide/adapters/claude-code-adapter.d.ts

@@ -1,18 +1,18 @@
 import type { IdeAdapter } from '../ide-types.js';
 /**
- * Claude Code adapter —— peaks-loop 的"起源 IDE"。
+ * Model-aware context-window size in tokens.
  *
- * 该 adapter 从原 `src/services/skills/hooks-settings-service.ts` 提取,保持
- * 字节级兼容:用户在 Claude Code 环境下跑 `peaks hooks install` 产出的
- * `.claude/settings.json` 与 refactor 前逐字节相同。
+ * Detection rule (documented):
+ *   1. Empty / unknown model → DEFAULT_CONTEXT_WINDOW_TOKENS (200_000).
+ *   2. Suffix heuristic — a model id containing `1m` (case-insensitive) is
+ *      treated as 1M-context.
+ *   3. Explicit allowlist — known 1M Claude model ids
+ *      (ONE_MILLION_CONTEXT_MODELS).
+ *   4. Everything else → 200_000 (the safe default).
  *
- * 字段解释(见 PRD AC-1):
- *   - dirName = '.claude'           : Claude Code 项目根下的 settings 目录
- *   - settingsFileName = 'settings.json'
- *   - envVar = 'CLAUDE_PROJECT_DIR' : Claude Code 注入的 env 变量,用于 ${...} 占位
- *   - hookEvent = 'PreToolUse'      : Claude Code hook 数组 key
- *   - toolMatcher = 'Bash' | 'Task' : PreToolUse 数组元素的 matcher 字段
- *
- * 不可消除的 per-IDE 字段(见 tech-doc.md §1.3)。
+ * Callers MAY additionally infer ≥1M from the observed token count: if
+ * `contextTokens > DEFAULT_CONTEXT_WINDOW_TOKENS`, the model cannot be a
+ * 200K model and must be ≥1M (see `readClaudeTranscriptEstimate`).
  */
+export declare function modelContextWindowTokens(model: string): number;
 export declare const CLAUDE_CODE_ADAPTER: IdeAdapter;