agent-working-memory 0.13.1 → 0.14.1

package/dist/mcp.js

@@ -80,6 +80,7 @@ import { startLoopLagMonitor } from './core/write-telemetry.js';
 import { buildWhoami, formatWhoami } from './core/whoami.js';
 import { renderTaskEndInvitation, validateRecipeWrite, recipeSlug, getRecipe } from './recipes/index.js';
 import { formatRecallResultLine } from './core/format-recall.js';
+import { packRecallByBudget, formatTokenFooter, estimateTokens } from './core/token-budget.js';
 // --- Incognito Mode ---
 // When AWM_INCOGNITO=1, register zero tools. Claude won't see memory tools at all.
 // No DB, no engines, no sidecar — just a bare MCP server that exposes nothing.
@@ -196,15 +197,15 @@ else {
         return 'unclassified';
     }
     // --- Tools ---
-    server.tool('memory_write', `Store a memory. The salience filter decides whether it's worth keeping (active), needs more evidence (staging), or should be discarded.
-
-CALL THIS PROACTIVELY — do not wait to be asked. Write memories when you:
-- Discover something about the codebase, bugs, or architecture
-- Make a decision and want to remember why
-- Encounter and resolve an error
-- Learn a user preference or project pattern
-- Complete a significant piece of work
-
+    server.tool('memory_write', `Store a memory. The salience filter decides whether it's worth keeping (active), needs more evidence (staging), or should be discarded.
+
+CALL THIS PROACTIVELY — do not wait to be asked. Write memories when you:
+- Discover something about the codebase, bugs, or architecture
+- Make a decision and want to remember why
+- Encounter and resolve an error
+- Learn a user preference or project pattern
+- Complete a significant piece of work
+
 The concept should be a short label (3-8 words). The content should be the full detail.`, {
         concept: z.string().describe('Short label for this memory (3-8 words)'),
         content: z.string().describe('Full detail of what was learned'),
@@ -355,27 +356,28 @@ The concept should be a short label (3-8 words). The content should be the full
                 }],
         };
     });
-    server.tool('memory_recall', `Recall memories relevant to a query. Uses cognitive activation — not keyword search.
-
-ALWAYS call this when:
-- Starting work on a project or topic (recall what you know)
-- Debugging (recall similar errors and solutions)
-- Making decisions (recall past decisions and outcomes)
-- The user mentions a topic you might have stored memories about
-
-Accepts either "query" or "context" parameter — both work identically.
+    server.tool('memory_recall', `Recall memories relevant to a query. Uses cognitive activation — not keyword search.
+
+ALWAYS call this when:
+- Starting work on a project or topic (recall what you know)
+- Debugging (recall similar errors and solutions)
+- Making decisions (recall past decisions and outcomes)
+- The user mentions a topic you might have stored memories about
+
+Accepts either "query" or "context" parameter — both work identically.
 Returns the most relevant memories ranked by text relevance, temporal recency, and associative strength.`, {
         query: z.string().optional().describe('What to search for — describe the situation, question, or topic'),
         context: z.string().optional().describe('Alias for query (either works)'),
-        limit: z.number().optional().default(5).describe('Max memories to return (default 5)'),
+        limit: z.number().optional().default(3).describe('Max memories to return (default 3). Measured on 400 real-store probes: k=3 is the only token-POSITIVE setting (+115 tok/recall, against -680 at k=5) and is 41% faster at p50, because `limit` sizes the cross-encoder rerank pool and rerank is ~90% of warm recall time. It costs ~0.5pp success@1 versus k=5 — about 3 queries in 400, inside noise. Raise it explicitly when you want breadth for orientation rather than a specific fact; that is the case this default deliberately does not optimise for.'),
         min_score: z.number().optional().default(0.05).describe('Minimum relevance score (default 0.05)'),
         include_staging: z.boolean().optional().default(false).describe('Include weak/unconfirmed memories?'),
         use_reranker: z.boolean().optional().default(true).describe('Use cross-encoder re-ranking for better relevance (default true)'),
         use_expansion: z.boolean().optional().default(true).describe('Expand query with synonyms for better recall (default true)'),
         memory_type: z.enum(['episodic', 'semantic', 'procedural']).optional().describe('Filter by memory type (omit to search all types)'),
         workspace: z.string().optional().describe('Search across all agents in this workspace (hive mode). Omit for agent-scoped recall only.'),
-        require_confidence: z.number().optional().describe('Opt-in: abstain (return []) when recall confidence is below this threshold. Typical values: 0.10 (strict), 0.25 (balanced), 0.40 (aggressive). Confidence is the shape of the result-score distribution; low confidence indicates a noisy or best-of-bad-bunch recall.'),
+        require_confidence: z.number().optional().default(0.05).describe('Abstain (return nothing) when recall confidence is below this threshold. Defaults to 0.05 — a LIGHT filter, chosen from measurement: it halves the rate of answering off-topic queries at zero cost to hit rate. Raise it only for push-style use where nobody asked (0.25 is what the prime hook uses). Do NOT raise it for ordinary recall: a miss is expensive, because the agent then reads the codebase instead (~2,106 tokens), so aggressive thresholds measurably destroy value — 0.20+ cut net tokens saved by 25% in tests/abstention-eval. Pass 0 to disable. NOTE: this is NOT min_score — it gates on the SHAPE of the score distribution across the whole result set, not per-result relevance, so memories that pass min_score can still be withheld. When that happens the reply says RECALL ABSTAINED and reports how many were withheld; an empty result is only absence when it does not.'),
         granularity: z.enum(['full', 'compact', 'auto']).optional().describe('Output granularity (Paper 3: cognitive teaming). "full" (default): no change. "compact": every result carries a short summary field. "auto": confidence-adaptive — top result gets a longer summary when there is a clear winner, otherwise everything is compact for scanning.'),
+        max_tokens: z.number().optional().describe('Token budget for the response. `limit` is a COUNT and is token-blind — 5 results may cost 400 tokens or 4,000. Use this when context is tight: results are packed by value-per-token until the budget is reached, the top-scored match always gets first refusal, and the reply reports what it cost and what was withheld. Omit for no budget (everything is returned, still with accounting).'),
     }, async (params) => {
         const queryText = params.query ?? params.context;
         if (!queryText) {
@@ -388,6 +390,9 @@ Returns the most relevant memories ranked by text relevance, temporal recency, a
         }
         // Use workspace from param, env var, or omit for agent-scoped
         const workspace = params.workspace ?? process.env.AWM_WORKSPACE ?? undefined;
+        // Set only when a gate WITHHELD results. Distinguishes abstention from absence,
+        // which an empty array cannot.
+        let abstained;
         const results = await activationEngine.activate({
             agentId: AGENT_ID,
             context: queryText,
@@ -400,6 +405,7 @@ Returns the most relevant memories ranked by text relevance, temporal recency, a
             workspace,
             requireConfidence: params.require_confidence,
             granularity: params.granularity,
+            onAbstain: (info) => { abstained = info; },
         });
         // Auto-checkpoint: track recall
         try {
@@ -413,27 +419,69 @@ Returns the most relevant memories ranked by text relevance, temporal recency, a
             ? formatPeerDecisions(queryPeerDecisions(coordDb, AGENT_ID, queryText))
             : '';
         if (results.length === 0) {
+            // Two very different conditions used to render identically as
+            // "No relevant memories found." — a claim of absence the system cannot make
+            // when a gate withheld matches. Callers read it as absence and stopped
+            // looking, for memories scoring well above minScore.
+            const text = abstained
+                ? [
+                    `RECALL ABSTAINED — this is NOT "no memories exist".`,
+                    ``,
+                    `${abstained.candidates} candidate${abstained.candidates === 1 ? '' : 's'} matched ` +
+                        `(best score ${abstained.topScore.toFixed(3)}) and ${abstained.candidates === 1 ? 'was' : 'were'} ` +
+                        `withheld because recall confidence ` +
+                        `${abstained.confidence !== undefined ? abstained.confidence.toFixed(3) + ' ' : ''}` +
+                        `fell below your require_confidence of ${abstained.threshold ?? '?'}.`,
+                    ``,
+                    `require_confidence is NOT min_score. It gates on the SHAPE of the score` +
+                        ` distribution across the whole result set, not on how relevant any single` +
+                        ` memory is — so results that comfortably pass min_score can still be withheld here.`,
+                    ``,
+                    `To see them: re-run this query with require_confidence: 0.`,
+                    `Do not conclude the memories are absent without doing that.`,
+                ].join('\n') + peerSuffix
+                : 'No relevant memories found.' + peerSuffix;
+            return { content: [{ type: 'text', text }] };
+        }
+        // Confidence-adaptive output (Paper 3: cognitive teaming) and D8
+        // (2026-07-30) conflict surfacing both live in the shared formatter now —
+        // see core/format-recall.ts for why it's extracted (0.12.1: unit-testable
+        // without booting the server) and why the id sits after the score.
+        // 0.13.3: pack to a token budget (if given) and always report the spend.
+        // `limit` is a count and is token-blind; max_tokens bounds what this call
+        // actually costs the caller's context. See core/token-budget.ts for why the
+        // top-scored result gets first refusal rather than pure density packing.
+        // Reserve what the rest of the reply will cost, so max_tokens bounds the
+        // WHOLE response and not just the result lines. FOOTER_ALLOWANCE is a
+        // fixed upper bound on the accounting line (measured at ~15-25 tokens; the
+        // unit test asserts it stays under 40).
+        const FOOTER_ALLOWANCE = 40;
+        const reserved = estimateTokens(peerSuffix) + FOOTER_ALLOWANCE;
+        const packed = packRecallByBudget(results, formatRecallResultLine, params.max_tokens, reserved);
+        // A budget too small to admit even the best match must not look like "no
+        // memories found" — that is a different fact and would send the caller off
+        // to read code that AWM could have answered. Say what actually happened so
+        // the fix (raise the budget) is obvious.
+        if (packed.kept === 0 && packed.total > 0) {
             return {
                 content: [{
                         type: 'text',
-                        text: 'No relevant memories found.' + peerSuffix,
+                        text: `${packed.total} memories matched, but none fit a ${params.max_tokens}-token budget `
+                            + `(smallest result is ~${packed.withheldTokens} tok across ${packed.total}). `
+                            + `Raise max_tokens, or use granularity: 'compact' to shrink each result.`
+                            + peerSuffix,
                     }],
             };
         }
-        // Confidence-adaptive output (Paper 3: cognitive teaming) and D8
-        // (2026-07-30) conflict surfacing both live in the shared formatter now —
-        // see core/format-recall.ts for why it's extracted (0.12.1: unit-testable
-        // without booting the server) and why the id sits after the score.
-        const lines = results.map(formatRecallResultLine);
         return {
             content: [{
                     type: 'text',
-                    text: lines.join('\n') + peerSuffix,
+                    text: packed.lines.join('\n') + peerSuffix + formatTokenFooter(packed, params.max_tokens),
                 }],
         };
     });
-    server.tool('memory_feedback', `Report whether a recalled memory was actually useful. This updates the memory's confidence score — useful memories become stronger, useless ones weaken.
-
+    server.tool('memory_feedback', `Report whether a recalled memory was actually useful. This updates the memory's confidence score — useful memories become stronger, useless ones weaken.
+
 Always call this after using a recalled memory so the system learns what's valuable.`, {
         engram_id: z.string().describe('ID of the memory (from memory_recall results)'),
         useful: z.boolean().describe('Was this memory actually helpful?'),
@@ -456,8 +504,8 @@ Always call this after using a recalled memory so the system learns what's valua
                 }],
         };
     });
-    server.tool('memory_retract', `Retract a memory that turned out to be wrong. Creates a correction and reduces confidence of related memories.
-
+    server.tool('memory_retract', `Retract a memory that turned out to be wrong. Creates a correction and reduces confidence of related memories.
+
 Use this when you discover a memory contains incorrect information.`, {
         engram_id: z.string().describe('ID of the wrong memory'),
         reason: z.string().describe('Why is this memory wrong?'),
@@ -481,13 +529,13 @@ Use this when you discover a memory contains incorrect information.`, {
                 }],
         };
     });
-    server.tool('memory_supersede', `Replace an outdated memory with a newer one. Unlike retraction (which marks memories as wrong), supersession marks the old memory as outdated but historically correct.
-
-Use this when:
-- A status or count has changed (e.g., "5 reviews done" → "7 reviews done")
-- Architecture or infrastructure evolved (e.g., "two-repo model" → "three-repo model")
-- A schedule or plan was updated
-
+    server.tool('memory_supersede', `Replace an outdated memory with a newer one. Unlike retraction (which marks memories as wrong), supersession marks the old memory as outdated but historically correct.
+
+Use this when:
+- A status or count has changed (e.g., "5 reviews done" → "7 reviews done")
+- Architecture or infrastructure evolved (e.g., "two-repo model" → "three-repo model")
+- A schedule or plan was updated
+
 The old memory stays in the database (searchable for history) but is heavily down-ranked in recall so the current version dominates.`, {
         old_engram_id: z.string().describe('ID of the outdated memory (from memory_recall results, or memory_write\'s own response if you just wrote it)'),
         new_engram_id: z.string().describe('ID of the replacement memory'),
@@ -518,7 +566,7 @@ The old memory stays in the database (searchable for history) but is heavily dow
         const info = await buildWhoami(store, AGENT_ID, 'mcp');
         return { content: [{ type: 'text', text: formatWhoami(info) }] };
     });
-    server.tool('memory_stats', `Get memory health stats — how many memories, confidence levels, association count, and system performance.
+    server.tool('memory_stats', `Get memory health stats — how many memories, confidence levels, association count, and system performance.
 Also shows the activity log path so the user can tail it to see what's happening.`, {}, async () => {
         const metrics = await evalEngine.computeMetrics(AGENT_ID);
         const checkpoint = await store.getCheckpoint(AGENT_ID);
@@ -549,13 +597,13 @@ Also shows the activity log path so the user can tail it to see what's happening
         };
     });
     // --- Checkpointing Tools ---
-    server.tool('memory_checkpoint', `Save your current execution state so you can recover after context compaction.
-
-ALWAYS call this before:
-- Long operations (multi-file generation, large refactors, overnight work)
-- Anything that might fill the context window
-- Switching to a different task
-
+    server.tool('memory_checkpoint', `Save your current execution state so you can recover after context compaction.
+
+ALWAYS call this before:
+- Long operations (multi-file generation, large refactors, overnight work)
+- Anything that might fill the context window
+- Switching to a different task
+
 Also call periodically during long sessions to avoid losing state. The state is saved per-agent and overwrites any previous checkpoint.`, {
         current_task: z.string().describe('What you are currently working on'),
         decisions: z.array(z.string()).optional().default([])
@@ -589,14 +637,14 @@ Also call periodically during long sessions to avoid losing state. The state is
                 }],
         };
     });
-    server.tool('memory_restore', `Restore your previous execution state after context compaction or at session start.
-
-Returns:
-- Your saved execution state (task, decisions, next steps, files)
-- Recently recalled memories for context
-- Your last write for continuity
-- How long you were idle
-
+    server.tool('memory_restore', `Restore your previous execution state after context compaction or at session start.
+
+Returns:
+- Your saved execution state (task, decisions, next steps, files)
+- Recently recalled memories for context
+- Your last write for continuity
+- How long you were idle
+
 Use this at the start of every session or after compaction to pick up where you left off.`, {}, async () => {
         const checkpoint = await store.getCheckpoint(AGENT_ID);
         // Cold-store nudge: an empty store means the agent has nothing to recall — offer to warm-start.
@@ -630,7 +678,11 @@ Use this at the start of every session or after compaction to pick up where you
                 const results = await activationEngine.activate({
                     agentId: AGENT_ID,
                     context: recallContext,
-                    limit: 5,
+                    // 3, not 5 — measured token-positive (+115/recall vs -680) and 41% faster.
+                    // The case is stronger here than for explicit recall: this is push-style, so
+                    // nobody asked for these results and every one of them costs context the user
+                    // did not request.
+                    limit: 3,
                     minScore: 0.05,
                     useReranker: true,
                     useExpansion: true,
@@ -715,9 +767,9 @@ Use this at the start of every session or after compaction to pick up where you
         if (coordDb) {
             try {
                 const myAgent = coordDb.prepare(`SELECT id FROM coord_agents WHERE name = ? AND status != 'dead' ORDER BY last_seen DESC LIMIT 1`).get(AGENT_ID);
-                const peerDecisions = coordDb.prepare(`SELECT d.summary, a.name AS author_name, d.created_at
-           FROM coord_decisions d JOIN coord_agents a ON d.author_id = a.id
-           WHERE d.author_id != ? AND d.created_at > datetime('now', '-30 minutes')
+                const peerDecisions = coordDb.prepare(`SELECT d.summary, a.name AS author_name, d.created_at
+           FROM coord_decisions d JOIN coord_agents a ON d.author_id = a.id
+           WHERE d.author_id != ? AND d.created_at > datetime('now', '-30 minutes')
            ORDER BY d.created_at DESC LIMIT 10`).all(myAgent?.id ?? '');
                 if (peerDecisions.length > 0) {
                     parts.push(`\n**Peer decisions (last 30 min):**`);
@@ -736,11 +788,11 @@ Use this at the start of every session or after compaction to pick up where you
         };
     });
     // --- Onboarding Tools (warm-start a cold store) ---
-    server.tool('onboard_scan', `Scan a project's documentation + repository and return CANDIDATE memories to seed a cold store.
-
-Use this when the store is empty / you're new to a project. The scan is deterministic
-(real file contents, not guesses) — YOUR job is to refine the candidates into atomic,
-recall-shaped memories, run the interview (onboard_questions), confirm with the user, then
+    server.tool('onboard_scan', `Scan a project's documentation + repository and return CANDIDATE memories to seed a cold store.
+
+Use this when the store is empty / you're new to a project. The scan is deterministic
+(real file contents, not guesses) — YOUR job is to refine the candidates into atomic,
+recall-shaped memories, run the interview (onboard_questions), confirm with the user, then
 save the good ones with memory_write (memory_class="canonical"). Nothing is saved by this tool.`, {
         docs: z.array(z.string()).optional()
             .describe('Doc files/dirs to scan (Markdown/text). Defaults to the repo (or cwd).'),
@@ -765,18 +817,18 @@ save the good ones with memory_write (memory_class="canonical"). Nothing is save
         ].join('\n');
         return { content: [{ type: 'text', text }] };
     });
-    server.tool('onboard_questions', `Return the onboarding interview questions. Ask the user ONE at a time, starting with the
+    server.tool('onboard_questions', `Return the onboarding interview questions. Ask the user ONE at a time, starting with the
 goal of the memory system, and ask follow-ups for clarity. Turn each answer into a canonical memory.`, {}, async () => ({
         content: [{ type: 'text', text: INTERVIEW_QUESTIONS.map((q, i) => `${i + 1}. ${q}`).join('\n') }],
     }));
     // --- Task Management Tools ---
-    server.tool('memory_task_add', `Create a task that you need to come back to. Tasks are memories with status and priority tracking.
-
-Use this when:
-- You identify work that needs doing but can't do it right now
-- The user mentions something to do later
-- You want to park a sub-task while focusing on something more urgent
-
+    server.tool('memory_task_add', `Create a task that you need to come back to. Tasks are memories with status and priority tracking.
+
+Use this when:
+- You identify work that needs doing but can't do it right now
+- The user mentions something to do later
+- You want to park a sub-task while focusing on something more urgent
+
 Tasks automatically get high salience so they won't be discarded.`, {
         concept: z.string().describe('Short task title (3-10 words)'),
         content: z.string().describe('Full task description — what needs doing, context, acceptance criteria'),
@@ -816,11 +868,11 @@ Tasks automatically get high salience so they won't be discarded.`, {
                 }],
         };
     });
-    server.tool('memory_task_update', `Update a task's status or priority. Use this to:
-- Start working on a task (open → in_progress)
-- Mark a task done (→ done)
-- Block a task on another (→ blocked)
-- Reprioritize (change priority)
+    server.tool('memory_task_update', `Update a task's status or priority. Use this to:
+- Start working on a task (open → in_progress)
+- Mark a task done (→ done)
+- Block a task on another (→ blocked)
+- Reprioritize (change priority)
 - Unblock a task (clear blocked_by)`, {
         task_id: z.string().describe('ID of the task to update'),
         status: z.enum(['open', 'in_progress', 'blocked', 'done']).optional()
@@ -850,8 +902,8 @@ Tasks automatically get high salience so they won't be discarded.`, {
                 }],
         };
     });
-    server.tool('memory_task_list', `List tasks with optional status filter. Shows tasks ordered by priority (urgent first).
-
+    server.tool('memory_task_list', `List tasks with optional status filter. Shows tasks ordered by priority (urgent first).
+
 Use at the start of a session to see what's pending, or to check blocked/done tasks.`, {
         status: z.enum(['open', 'in_progress', 'blocked', 'done']).optional()
             .describe('Filter by status (omit to see all active tasks)'),
@@ -877,10 +929,10 @@ Use at the start of a session to see what's pending, or to check blocked/done ta
                 }],
         };
     });
-    server.tool('memory_task_next', `Get the single most important task to work on next.
-
-Prioritizes: in_progress tasks first (finish what you started), then by priority level, then oldest first. Skips blocked and done tasks.
-
+    server.tool('memory_task_next', `Get the single most important task to work on next.
+
+Prioritizes: in_progress tasks first (finish what you started), then by priority level, then oldest first. Skips blocked and done tasks.
+
 Use this when you finish a task or need to decide what to do next.`, {}, async () => {
         const next = await store.getNextTask(AGENT_ID);
         if (!next) {
@@ -896,13 +948,13 @@ Use this when you finish a task or need to decide what to do next.`, {}, async (
         };
     });
     // --- Task Bracket Tools ---
-    server.tool('memory_task_begin', `Signal that you're starting a significant task. Auto-checkpoints current state and recalls relevant memories.
-
-CALL THIS when starting:
-- A multi-step operation (doc generation, large refactor, migration)
-- Work on a new topic or project area
-- Anything that might fill the context window
-
+    server.tool('memory_task_begin', `Signal that you're starting a significant task. Auto-checkpoints current state and recalls relevant memories.
+
+CALL THIS when starting:
+- A multi-step operation (doc generation, large refactor, migration)
+- Work on a new topic or project area
+- Anything that might fill the context window
+
 This ensures your state is saved before you start, and primes recall with relevant context.`, {
         topic: z.string().describe('What task are you starting? (3-15 words)'),
         files: z.array(z.string()).optional().default([])
@@ -928,7 +980,8 @@ This ensures your state is saved before you start, and primes recall with releva
             const results = await activationEngine.activate({
                 agentId: AGENT_ID,
                 context: params.topic,
-                limit: 5,
+                // 3, not 5 — see memory_recall. Push-style auto-recall on task start.
+                limit: 3,
                 minScore: 0.05,
                 useReranker: true,
                 useExpansion: true,
@@ -953,13 +1006,13 @@ This ensures your state is saved before you start, and primes recall with releva
                 }],
         };
     });
-    server.tool('memory_task_end', `Signal that you've finished a significant task. Writes a summary memory and auto-checkpoints.
-
-CALL THIS when you finish:
-- A multi-step operation
-- Before switching to a different topic
-- At the end of a work session
-
+    server.tool('memory_task_end', `Signal that you've finished a significant task. Writes a summary memory and auto-checkpoints.
+
+CALL THIS when you finish:
+- A multi-step operation
+- Before switching to a different topic
+- At the end of a work session
+
 This captures what was accomplished so future sessions can recall it.`, {
         summary: z.string().describe('What was accomplished? Include key outcomes, decisions, and any issues.'),
         tags: z.array(z.string()).optional().default([])
@@ -1035,14 +1088,14 @@ This captures what was accomplished so future sessions can recall it.`, {
                 }],
         };
     });
-    server.tool('compress_output', `Compress a STRUCTURED tool output (JSON object/array, query rows, log records) into TOON —
-a compact, schema-aware tabular encoding — before putting it in your context. Cuts ~50-65%
-of the tokens on uniform arrays at zero comprehension cost (validated: models read TOON as
-accurately as JSON). Use this on large tool results you need to keep in context.
-
-Output-only and safe: it never changes the data. Non-JSON / prose is returned unchanged.
-TOON is only emitted when it reproduces the input exactly (self-verified round-trip);
-otherwise you get plain JSON back. When compressed, you also get a 'ref' — call
+    server.tool('compress_output', `Compress a STRUCTURED tool output (JSON object/array, query rows, log records) into TOON —
+a compact, schema-aware tabular encoding — before putting it in your context. Cuts ~50-65%
+of the tokens on uniform arrays at zero comprehension cost (validated: models read TOON as
+accurately as JSON). Use this on large tool results you need to keep in context.
+
+Output-only and safe: it never changes the data. Non-JSON / prose is returned unchanged.
+TOON is only emitted when it reproduces the input exactly (self-verified round-trip);
+otherwise you get plain JSON back. When compressed, you also get a 'ref' — call
 retrieve_original(ref) to get the verbatim source back if you ever need it.`, {
         output: z.string().describe('The tool output to compress — JSON text (preferred) or any string. Non-JSON is returned unchanged.'),
         min_saving_chars: z.number().optional().describe('Only emit TOON if it saves at least this many characters (default 40).'),
@@ -1056,8 +1109,8 @@ retrieve_original(ref) to get the verbatim source back if you ever need it.`, {
             content: [{ type: 'text', text: header + r.text }],
         };
     });
-    server.tool('retrieve_original', `Retrieve the verbatim original text for a 'ref' returned by compress_output. Use this when
-you need the exact, uncompressed source (e.g. to pass it to another tool unchanged). Returns
+    server.tool('retrieve_original', `Retrieve the verbatim original text for a 'ref' returned by compress_output. Use this when
+you need the exact, uncompressed source (e.g. to pass it to another tool unchanged). Returns
 an error if the ref has expired (originals are kept for the most recent compressions only).`, {
         ref: z.string().describe('The ref handle returned by compress_output (e.g. "awm_orig_12").'),
     }, async (params) => {