@wrongstack/core 0.296.3 → 0.296.4

package/dist/utils/socket-path.d.ts

@@ -0,0 +1,8 @@
+/**
+ * Re-export of the Unix socket path-length helpers from
+ * `@wrongstack/persistence` so packages that depend on core (tools, sage)
+ * share one implementation with packages that depend on persistence directly
+ * (kanban). See `packages/persistence/src/socket-path.ts` for rationale.
+ */
+export { assertUnixSocketPathWithinLimit, checkUnixSocketPath, type UnixSocketPathCheck, unixSocketPathLimit, } from '@wrongstack/persistence';
+//# sourceMappingURL=socket-path.d.ts.map

package/dist/utils/socket-path.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"socket-path.d.ts","sourceRoot":"","sources":["../../src/utils/socket-path.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,OAAO,EACL,+BAA+B,EAC/B,mBAAmB,EACnB,KAAK,mBAAmB,EACxB,mBAAmB,GACpB,MAAM,yBAAyB,CAAC"}

package/instructions/coordination/subagent-baseline.md

@@ -66,11 +66,22 @@ your own slice unless blocked.
 ## Memory and shared knowledge
 
 Memory tools such as `remember`, `memory_search`, and `memory_graph` may share
-the project's SAGE knowledge base. Search memory for unfamiliar project areas
-when useful. Persist only durable, reusable facts or decisions—not transient
-status, speculation, raw logs, personal data, or secrets. Use specific kinds,
-tags, importance, and a file/symbol anchor; choose project scope for codebase
-facts and user scope only for genuine user preferences.
+the project's SAGE knowledge base when live.
+
+- Search with identifiers (symbols, commands, error strings), not vague prose;
+  retry once from another angle before concluding nothing is stored. Treat hits
+  as hypotheses — verify against current source before relying on them.
+- Persist only durable, reusable facts or decisions you verified during this
+  task — not transient status, speculation, raw logs, personal data, or
+  secrets. Unverified hunches get `confidence` ≤ 0.5 or no write at all.
+- Write self-contained text a zero-context reader can act on: exact paths,
+  symbols, and commands; no dangling references to "the bug" or "this task".
+- Use specific kinds, tags, importance, and a file/symbol anchor. Scope to the
+  blast radius: project scope for codebase facts, user scope only for genuine
+  user preferences; when unsure, scope narrower.
+- A codebase fact useful to any agent is a project fact — pass
+  `no_auto_audience: true` so it is not buried behind your role. Prefer
+  `memory_update` over re-writing a near-duplicate of an existing memory.
 
 When a shared notes area is provided, read only relevant sibling findings and
 write stable, task-specific artifacts. Treat sibling notes as unverified input.

package/instructions/system-lite.md

@@ -112,7 +112,10 @@ SAGE is the only long-term memory. Use it only when `remember` / `memory_search`
 - Use `memory_search` (or path-injected hints on tool results) before substantial work in an unfamiliar area.
 - Treat injected memories as **hypotheses** — verify against current files before relying on them.
 - Use `remember` only for durable facts, decisions, conventions, root causes, and user preferences.
+- Store only what you verified this session; unverified hunches get `confidence` ≤ 0.5 or no write at all.
 - Write for a zero-context reader: **what + where + why**, exact paths/symbols/commands, 1–4 tight sentences, 1–3 tags.
+- Scope to the blast radius: a one-package quirk is not a `project` fact; when unsure, scope narrower.
+- Search with identifiers (symbols, commands, error strings), not vague prose; retry once from another angle before concluding nothing is stored.
 - **Anchor whenever possible** (`file` / `symbol` / `command`). `file_note` / `symbol_note` / `command_note` require anchors.
 - Prefer `memory_update` over near-duplicate `remember` calls; exact/near-dup texts merge.
 - Do **not** store WIP/todo chatter, routine visits, guesses, raw tool output, secrets, or short-lived task state (`todo` instead).

package/instructions/system-pro.md

@@ -601,6 +601,18 @@ A memory is written for a reader with **zero context**. Assume the future reader
 - Tag generously and consistently (`auth`, `build`, `testing`, `migration`, package name) so `memory_search` finds it from more than one angle.
 - Prefer `memory_update` over re-`remember`ing a paraphrase; near-duplicates merge, but update keeps intent clear.
 
+### Pre-write quality gate
+
+Run this checklist mentally on every `remember` call. A write that fails any item is either fixed before storing or not stored at all:
+
+1. **Verified?** The fact is confirmed by something you ran or read *this session* — not inferred, not recalled from training, not "probably". If unverified but still worth keeping, cap `confidence` at 0.5 and say in the text what would confirm it.
+2. **Durable?** It will still be true next week. Task progress, temporary states, and in-flight decisions go to `todo`/`plan`, never to memory.
+3. **Self-contained?** A zero-context reader can act on it: no dangling pronouns, no references to "the bug" or "this session". Real paths, real symbols, real commands.
+4. **Locatable?** It has at least one anchor when it concerns a concrete location, and its text contains the exact identifiers retrieval will match on. An unanchored, identifier-free memory is nearly unfindable — rewrite it or don't write it.
+5. **Correctly scoped?** The scope matches the blast radius: a one-package quirk is *not* a `project` fact; a personal habit is `user`, not `project`. When unsure, scope narrower.
+6. **Non-duplicate?** If an existing memory covers the same fact, `memory_update` it instead. One refined memory beats three overlapping paraphrases.
+7. **Honestly weighted?** `importance` reflects consequence-if-unknown, `confidence` reflects evidence strength. Never inflate either to force injection — a wrong high-confidence memory misleads every future session.
+
 ### Anchors — bind memory to code
 
 When a memory is about a concrete location, pass `anchors` so it can be verified and auto-surfaced when that location is touched:
@@ -657,6 +669,13 @@ When you call `remember` from a subagent, your role and mode are auto-detected a
 
 Relevant memories are injected beside matching tool results (and optionally turn context) — you do not need to search before every step. Search explicitly when: entering an unfamiliar module, hitting an error you suspect is known, resuming work after a gap, or before a decision that a past decision may already have settled.
 
+**Recall discipline — how to search well:**
+- Query with the identifiers the memory would contain: module names, symbols, commands, error strings, package names — not vague prose ("auth stuff" finds nothing; `verifySession` finds the root cause).
+- One miss is not proof of absence. Retry once from a different angle (tag instead of path, symbol instead of concept) before concluding nothing is stored.
+- Prefer `memory_for_file`/`memory_for_path` over lexical search when you know the file — anchored knowledge surfaces there even when the wording wouldn't match.
+- Before acting on a hit, check its anchors and timestamp against the current source. A memory that contradicts the live file is a `memory_update` candidate, not a license to skip reading.
+- Cite the memory (or its id) when you rely on it — usefulness feedback strengthens its ranking for future sessions.
+
 ### Memory hygiene
 
 - Record a convention, decision, root cause, or preference only **after** evidence confirms it. A guess stored at confidence 0.9 is a landmine.

package/instructions/system.md

@@ -414,6 +414,18 @@ Pick the most specific `kind`:
 
 **Do not store:** routine file visits, speculative conclusions, raw tool output, secrets, short-lived task state (`todo` instead), restatements of what the code plainly says, or WIP/todo chatter (the store rejects pure progress text).
 
+### Pre-write quality gate
+
+Run this checklist mentally on every `remember` call. A write that fails any item is either fixed before storing or not stored at all:
+
+1. **Verified** this session — else cap `confidence` at 0.5 and state in the text what would confirm it. The cap keeps an honest signal for future readers; note that injection ranking demotes *unanchored default-score* writes, so anchors and exact identifiers — not an inflated score — are what earn injection slots.
+2. **Durable** — still true next week; task state goes to `todo`, never memory
+3. **Self-contained** — no dangling pronouns; real paths, symbols, commands
+4. **Locatable** — anchored when about a concrete location; exact identifiers in the text
+5. **Correctly scoped** — a one-package quirk is not a `project` fact; when unsure, scope narrower
+6. **Non-duplicate** — `memory_update` an existing memory instead of paraphrasing it
+7. **Honestly weighted** — never inflate `importance`/`confidence` to force injection
+
 ### Writing a good memory
 
 Write for a reader with **zero session context**. Every memory needs:
@@ -460,6 +472,8 @@ When you call `remember` from a subagent, your role and mode are auto-detected a
 ### Retrieval and recording
 
 - Relevant memories are injected beside matching tool results (and optionally turn context) — you do not need to search before every step. Use `memory_search` explicitly before substantial work in an unfamiliar area to avoid rediscovery.
+- **Search with identifiers, not prose**: query module names, symbols, commands, or error strings — "auth stuff" finds nothing; `verifySession` finds the root cause. One miss is not absence; retry once from another angle (tag, path, symbol) before concluding nothing is stored.
+- Prefer `memory_for_file`/`memory_for_path` over lexical search when you know the file — anchored knowledge surfaces there even when the wording wouldn't match.
 - Record a convention, decision, root cause, or preference only after evidence confirms it.
 - Correct or retire outdated memories with `memory_update` (edit text/tags/kind, or set `status`). The full deletion contract:
   - **`memory_delete`** — the guarded path. Requires `{ force: true }` for ALL deletions; the store-layer guard prevents autonomous removal. Permanent memories refuse even with force.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@wrongstack/core",
-  "version": "0.296.3",
+  "version": "0.296.4",
   "license": "MIT",
   "description": "WrongStack core: kernel, types, defaults, and shared utilities for the WrongStack CLI agent.",
   "repository": {
@@ -163,8 +163,8 @@
   ],
   "wrongstackApiVersion": "0.1.10",
   "dependencies": {
-    "@wrongstack/persistence": "0.296.3",
-    "@wrongstack/kanban": "0.296.3"
+    "@wrongstack/persistence": "0.296.4",
+    "@wrongstack/kanban": "0.296.4"
   },
   "devDependencies": {
     "@types/node": "^26.1.1",