@wrongstack/core 0.291.1 → 0.292.0

package/dist/utils/token-estimate.d.ts

@@ -32,6 +32,20 @@ export declare function computeMessageTokens(msg: Message): number;
  * sum for fully-cached arrays.
  */
 export declare function estimateMessageTokens(messages: readonly Message[]): number;
+/**
+ * Real-usage-anchored input-token count. Given the provider's authoritative
+ * prompt-token count from the last response (`anchorTokens`, a REAL number) and
+ * the `messages.length` of the request that produced it (`anchorMsgCount`),
+ * returns `anchorTokens + estimate(messages appended since)` — so everything up
+ * to the last turn is exact and only the newest, not-yet-sent messages are
+ * estimated. This is deliberately NOT calibrated: the base is already real, and
+ * on the next response the whole thing re-anchors to the new real count.
+ *
+ * Returns `null` when there is no usable anchor (no response yet, or the
+ * message array shrank below the anchor — e.g. after compaction — in which case
+ * the caller falls back to a full estimate until the next response re-anchors).
+ */
+export declare function realAnchoredInputTokens(messages: readonly Message[], anchorTokens: number | undefined, anchorMsgCount: number | undefined): number | null;
 /**
  * Rough estimate of tokens in a tool definition (name + description + schema).
  * Accounts for the JSON-serialized inputSchema which is sent to the API
@@ -103,6 +117,19 @@ export declare function estimateRequestTokensCalibrated(messages: unknown, syste
     description?: string | undefined;
     inputSchema: unknown;
 }[], calibrationKey?: string): RequestTokenBreakdown;
+/**
+ * Never-undercount upper bound for the request token total, for the **send
+ * guard** only. Takes the flat estimate and scales it up by the greater of the
+ * content-density multiplier and the calibration ceiling, so the guarded value
+ * satisfies `real ≤ upperBound`. The context bar and `/context` keep using the
+ * calibrated estimate — this deliberately over-counts, which is only ever safe
+ * for the "must this be trimmed before sending?" decision.
+ */
+export declare function estimateRequestTokensUpperBound(messages: unknown, systemPrompt: unknown, tools: {
+    name: string;
+    description?: string | undefined;
+    inputSchema: unknown;
+}[], calibrationKey?: string): RequestTokenBreakdown;
 /**
  * Resets calibration state. Primarily for tests that run in the same
  * process and need a clean slate between suites. With no argument it clears

package/dist/utils/token-estimate.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"token-estimate.d.ts","sourceRoot":"","sources":["../../src/utils/token-estimate.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,OAAO,EAAE,MAAM,sBAAsB,CAAC;AAyGpD;;;;GAIG;AACH,wBAAgB,uBAAuB,CAAC,KAAK,EAAE,OAAO,GAAG,MAAM,CAQ9D;AAED;;GAEG;AACH,wBAAgB,wBAAwB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,GAAG,MAAM,CAG1E;AAED;;GAEG;AACH,wBAAgB,kBAAkB,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAEvD;AAED;;;;;GAKG;AACH,wBAAgB,oBAAoB,CAAC,GAAG,EAAE,OAAO,GAAG,MAAM,CAUzD;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,SAAS,OAAO,EAAE,GAAG,MAAM,CAU1E;AAED;;;;GAIG;AACH,wBAAgB,qBAAqB,CAAC,IAAI,EAAE;IAC1C,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;IACjC,WAAW,EAAE,OAAO,CAAC;CACtB,GAAG,MAAM,CAWT;AAED;;;;;;;;;;;GAWG;AACH,MAAM,WAAW,qBAAqB;IACpC,QAAQ,EAAE,MAAM,CAAC;IACjB,YAAY,EAAE,MAAM,CAAC;IACrB,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;CACf;AAED,wBAAgB,qBAAqB,CACnC,QAAQ,EAAE,OAAO,EACjB,YAAY,EAAE,OAAO,EACrB,KAAK,EAAE;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,WAAW,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;IAAC,WAAW,EAAE,OAAO,CAAA;CAAE,EAAE,EACjF,cAAc,GAAE,MAA+B,GAC9C,qBAAqB,CAqEvB;AAED;;;;;;;;;;;;;GAaG;AACH,wBAAgB,iBAAiB,CAC/B,iBAAiB,EAAE,MAAM,EACzB,oBAAoB,CAAC,EAAE,MAAM,EAC7B,cAAc,GAAE,MAA+B,GAC9C,IAAI,CAiBN;AAED;;;GAGG;AACH,wBAAgB,mBAAmB,CAAC,cAAc,GAAE,MAA+B,GAAG;IACpF,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd,UAAU,EAAE,OAAO,CAAC;CACrB,CAOA;AAED;;;;;;;;GAQG;AACH,wBAAgB,+BAA+B,CAC7C,QAAQ,EAAE,OAAO,EACjB,YAAY,EAAE,OAAO,EACrB,KAAK,EAAE;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,WAAW,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;IAAC,WAAW,EAAE,OAAO,CAAA;CAAE,EAAE,EACjF,cAAc,GAAE,MAA+B,GAC9C,qBAAqB,CA2BvB;AAWD;;;;GAIG;AACH,wBAAgB,gBAAgB,CAAC,cAAc,CAAC,EAAE,MAAM,GAAG,IAAI,CAM9D"}
+{"version":3,"file":"token-estimate.d.ts","sourceRoot":"","sources":["../../src/utils/token-estimate.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,OAAO,EAAE,MAAM,sBAAsB,CAAC;AAyGpD;;;;GAIG;AACH,wBAAgB,uBAAuB,CAAC,KAAK,EAAE,OAAO,GAAG,MAAM,CAQ9D;AAED;;GAEG;AACH,wBAAgB,wBAAwB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,GAAG,MAAM,CAG1E;AAED;;GAEG;AACH,wBAAgB,kBAAkB,CAAC,IAAI,EAAE,MAAM,GAAG,MAAM,CAEvD;AAED;;;;;GAKG;AACH,wBAAgB,oBAAoB,CAAC,GAAG,EAAE,OAAO,GAAG,MAAM,CAUzD;AAED;;;;;;;;;;GAUG;AACH,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,SAAS,OAAO,EAAE,GAAG,MAAM,CAU1E;AAED;;;;;;;;;;;;GAYG;AACH,wBAAgB,uBAAuB,CACrC,QAAQ,EAAE,SAAS,OAAO,EAAE,EAC5B,YAAY,EAAE,MAAM,GAAG,SAAS,EAChC,cAAc,EAAE,MAAM,GAAG,SAAS,GACjC,MAAM,GAAG,IAAI,CAMf;AAED;;;;GAIG;AACH,wBAAgB,qBAAqB,CAAC,IAAI,EAAE;IAC1C,IAAI,EAAE,MAAM,CAAC;IACb,WAAW,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;IACjC,WAAW,EAAE,OAAO,CAAC;CACtB,GAAG,MAAM,CAWT;AAED;;;;;;;;;;;GAWG;AACH,MAAM,WAAW,qBAAqB;IACpC,QAAQ,EAAE,MAAM,CAAC;IACjB,YAAY,EAAE,MAAM,CAAC;IACrB,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;CACf;AAED,wBAAgB,qBAAqB,CACnC,QAAQ,EAAE,OAAO,EACjB,YAAY,EAAE,OAAO,EACrB,KAAK,EAAE;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,WAAW,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;IAAC,WAAW,EAAE,OAAO,CAAA;CAAE,EAAE,EACjF,cAAc,GAAE,MAA+B,GAC9C,qBAAqB,CAqEvB;AAED;;;;;;;;;;;;;GAaG;AACH,wBAAgB,iBAAiB,CAC/B,iBAAiB,EAAE,MAAM,EACzB,oBAAoB,CAAC,EAAE,MAAM,EAC7B,cAAc,GAAE,MAA+B,GAC9C,IAAI,CAiBN;AAED;;;GAGG;AACH,wBAAgB,mBAAmB,CAAC,cAAc,GAAE,MAA+B,GAAG;IACpF,KAAK,EAAE,MAAM,CAAC;IACd,KAAK,EAAE,MAAM,CAAC;IACd,UAAU,EAAE,OAAO,CAAC;CACrB,CAOA;AAED;;;;;;;;GAQG;AACH,wBAAgB,+BAA+B,CAC7C,QAAQ,EAAE,OAAO,EACjB,YAAY,EAAE,OAAO,EACrB,KAAK,EAAE;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,WAAW,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;IAAC,WAAW,EAAE,OAAO,CAAA;CAAE,EAAE,EACjF,cAAc,GAAE,MAA+B,GAC9C,qBAAqB,CA2BvB;AA6DD;;;;;;;GAOG;AACH,wBAAgB,+BAA+B,CAC7C,QAAQ,EAAE,OAAO,EACjB,YAAY,EAAE,OAAO,EACrB,KAAK,EAAE;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,WAAW,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;IAAC,WAAW,EAAE,OAAO,CAAA;CAAE,EAAE,EACjF,cAAc,GAAE,MAA+B,GAC9C,qBAAqB,CAgBvB;AAWD;;;;GAIG;AACH,wBAAgB,gBAAgB,CAAC,cAAc,CAAC,EAAE,MAAM,GAAG,IAAI,CAM9D"}

package/instructions/llm/memory-consolidator.md

@@ -1,10 +1,13 @@
 You are a memory consolidator. Review the following session summary and decide what key facts, conventions, decisions, or learnings should be persisted to long-term memory.
 
 Session summary ({{iterations}} iterations):
-{{summary}}{{existingEntries}}
+{{summary}}
+
+Grounding evidence from this run (bounded; use only what is relevant):
+{{evidence}}{{existingEntries}}
 
 Return a JSON object with an "operations" array. Each operation must have an "action" field:
-- "add": create a new memory entry. Include "text", and optionally "type", "tags", "priority".
+- "add": create a new memory entry. Include "text", "type", "tags", "priority", "confidence", and concrete "anchors" when evidence supports them.
 
 This consolidator is strictly add-only. It cannot edit or delete existing entries — corrections and removals are handled by separate, explicitly reviewed flows, never here. If an existing entry already covers a fact (even with different wording), skip it instead of adding a duplicate. "edit" or "delete" operations will be ignored.
 
@@ -15,6 +18,12 @@ Memory types:
 - "preference": User or team preference (e.g. "prefers short variable names")
 - "reference": Pointer to a file or location (e.g. "auth logic in packages/core/src/auth/")
 - "anti_pattern": Something to avoid (e.g. "never use any in TypeScript")
+- "warning": A durable operational or safety warning
+- "workflow": A repeatable multi-step project workflow
+- "bug_root_cause": A verified cause of a recurring or important bug
+- "file_note": Durable ownership or responsibility of a file/package
+- "symbol_note": Durable behavior or contract of a function/class/symbol
+- "command_note": A useful project command and what it verifies or changes
 
 Priority levels:
 - "critical": Must always be known (e.g. security constraints)
@@ -25,10 +34,15 @@ Priority levels:
 Rules:
 - Only persist facts likely useful across multiple future sessions.
 - Do NOT persist task progress, temporary state, or one-off observations.
+- Prefer memories tied to stable project entities: files, directories, symbols, packages, tests, commands, decisions, conventions, workflows, and verified root causes.
+- Add 1-3 anchors whenever possible. Anchor objects use {"type":"file|directory|symbol|package|command|test|git", "path"?:string, "symbol"?:string, "command"?:string}.
+- Never invent an anchor. Use only paths, commands, symbols, packages, or tests supported by the summary/evidence.
+- Set confidence from 0.5 to 1.0. Use high confidence only for directly observed or verified claims.
 - Do NOT add a fact that an existing entry already covers.
 - Assign a type and priority to every "add" operation.
 - Use 1-3 hashtag tags for each entry (e.g. #typescript #build).
 - Be concise — each memory entry should be one clear sentence.
+- Return at most 5 additions. Return an empty operations array if the run produced no durable knowledge.
 
 Return ONLY valid JSON, no markdown, no explanation:
 {
@@ -38,7 +52,9 @@ Return ONLY valid JSON, no markdown, no explanation:
       "text": "Project uses pnpm workspaces with TypeScript strict mode",
       "type": "convention",
       "priority": "high",
-      "tags": ["pnpm", "typescript", "build"]
+      "confidence": 0.95,
+      "tags": ["pnpm", "typescript", "build"],
+      "anchors": [{"type":"file", "path":"pnpm-workspace.yaml"}]
     }
   ]
 }

package/instructions/system.md

@@ -87,8 +87,16 @@ I am composed of tool groups, each with a distinct purpose. This section maps th
 
 ### Agents & Delegation
 `delegate`, `spawn_subagent`, `assign_task`, `await_tasks`, `ask_subagent`, `terminate_subagent`, `fleet`, `fleet_emit`, `work_complete`, `quality_gate`, `collab_debug`
-- `delegate` for one-shot work in a separate context (own LLM, own budget).
-- `spawn_subagent` + `assign_task` + `await_tasks` for long-running fleet work.
+
+**The blocking-vs-async distinction is the most important rule in this section:**
+
+- `delegate` is **synchronous / blocking**: the leader's iteration pauses for the full duration of the subagent's run; no other tools execute while `delegate` is in flight. Use it only when your next decision genuinely needs the result (review, fact-check, sign-off). Multiple sequential `delegate` calls each block the leader, wasting wall-clock time.
+- `spawn_subagent` + `assign_task` + `await_tasks` is the **async / non-blocking** pattern: `spawn_subagent` returns immediately with a `subagentId`, `assign_task` returns immediately with a `taskId`, the leader keeps doing other work, and `await_tasks` retrieves the result later. Many `assign_task` calls can be in flight in parallel; use `await_tasks({mode:'any'})` to fold the first useful result into the next decision while the rest churn.
+
+**Decision rule:** does my next step depend on the result? If **yes** → `delegate`. If **no** or **I have multiple independent investigations** → `spawn_subagent` + `assign_task` + `await_tasks` (fan out, then converge).
+
+- `delegate` for one-shot work in a separate context (own LLM, own budget) — *blocking*.
+- `spawn_subagent` + `assign_task` + `await_tasks` for long-running fleet work — *non-blocking; the canonical async pattern*.
 - `quality_gate` to verify implementation before accepting it.
 - `collab_debug` for parallel bug-hunt / refactor / critique sessions.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@wrongstack/core",
-  "version": "0.291.1",
+  "version": "0.292.0",
   "license": "MIT",
   "description": "WrongStack core: kernel, types, defaults, and shared utilities for the WrongStack CLI agent.",
   "repository": {
@@ -91,7 +91,7 @@
   ],
   "wrongstackApiVersion": "0.1.10",
   "dependencies": {
-    "@wrongstack/kanban": "0.291.1"
+    "@wrongstack/kanban": "0.292.0"
   },
   "devDependencies": {
     "@types/node": "^26.1.1",