@try-works/dsh-memory 0.1.18 → 0.1.20

package/STATUS.md

@@ -1,287 +1,287 @@
-# dsh-memory — STATUS
-
-A tracker and planner: what exists, what is verified working, what is written but not
-loaded, and what is left. Kept as a record of MEASURED state, not intention.
-
-**Last measured: 2026-10-06** (347 tests passing, 41 src files, 50 test files).
-
-## How to read this
-
-Four states, and the difference between them is the whole point of the document:
-
-| mark | meaning |
-|---|---|
-| **LIVE** | verified working in the running host, by observing an effect it produced |
-| **WRITTEN** | implemented and tested on disk; the running host has not loaded it |
-| **NOT VERIFIED** | believed to work; no observation confirms it. Treat as unknown |
-| **TODO** | not done |
-
-This distinction exists because of two failures in this project. A retry loop was reported
-done and was logically correct but could never run, starved behind a backlog in a bounded
-batch. And session content search was reported enabled on the strength of a config flag; the
-database was created and the index was empty. In both cases a green result masked a dead
-feature. Both are recorded below as they actually are.
-
-## What dsh-memory is
-
-A DSH plugin storing durable agent memory as layered Markdown under a configurable root, with
-a global layer and per-workspace layers, committable to a git origin. Markdown is the record;
-indexes and state are derived and gitignored.
-
-| layer | scope | contents |
-|---|---|---|
-| `semantic` | global + workspace | facts, preferences, standing rules |
-| `episodic` | global + workspace | what happened, session by session |
-| `procedural` | global + workspace | how to do things here |
-| `learning` | global + workspace | lessons, usually from bugs |
-| `decision` | global + workspace | settled choices with their reasons |
-| `codebase` | workspace only | what each module does |
-| `history` | workspace only | commits, deduped by content |
-| `state.md` | workspace only | the repository now, replaced per section, freshness-marked |
-| `working.md` | workspace only | the agent now: task, activity, next steps |
-
-Live store: **1200 entries** — decision 264, learning 280, semantic 278, codebase 173,
-procedural 103, episodic 62, history 40.
-## Built and LIVE
-
-Observed in the running host. Each row names the effect that was seen, not a claim.
-
-| capability | evidence |
-|---|---|
-| Queue is the only state store | `state/retired/sessions.json`; `state/queue.jsonl` holds 170 entries |
-| Capture writes transitions | queue states move with every session; `summarized` 19 → 23 since boot |
-| Consolidation writes transitions | `consolidate` events fire; digests reach `added` |
-| Retry scheduling | one entry is `failed` with a reason; `due()` returns it as `retry` |
-| Stalled-stage watchdog | `reapStalled` runs before each batch; 0 stranded mid-stage |
-| Explicit abandonment | states drain rather than accumulate; `ABANDONED` reported by `problems()` |
-| `memory_status` reports the queue | per-state counts, parked reasons, grouped attention with ages |
-| Attention is GROUPED | 2 rows where the first version produced 79, 78 of them identical |
-| Injection selects by RELEVANCE | the context card reads "Relevant learnings" and lists this session's work, where it previously always listed UI notes |
-| History ingest | `history-check` events, 5 since the instrumented host started |
-| Injection metric counts injections | `events.inject` advances by 1 per injection; the old character total is preserved under `legacy.inject` |
-| `memory_recover` is wired | calling it no longer answers "unavailable in this context"; it reaches output validation |
-| Session logs are searchable in principle | zstd, not encrypted; 368 logs, 740MB compressed, ~2.96GB plaintext |
-
-## WRITTEN but NOT LOADED
-
-Implemented and tested on disk. The running host predates them, so none of this is in effect.
-
-| change | why it matters | how it was verified on disk |
-|---|---|---|
-| `due()` priority ordering | retry → abandon → work → recover, so a bounded batch reaches a due retry | measured: the retry sat at position 101 of 111 and is now position 1 |
-| `hostExtras(host)` as one source | `index.js` and the wiring test shared a copied list that drifted | the test now derives extras from the host object |
-| `memory_recover` report omits its `entries` field rather than blanking it | the old line added `entries: undefined`, which is NOT valid JSON, so the harness rejected the whole result | the new guard round-trips every tool output; reintroducing it fails with `value.entries is undefined` |
-
-How to tell loaded from not: ask the host. `memory_recover` used to reply *"unavailable in this
-context"* and now replies *"returned invalid output: value is not lossless JSON"*. The change of
-ERROR is the proof the first fix is loaded and the tool now executes — a different failure, one
-layer further along. The report fix in the last row is what remains pending.
-### The tasks layer (new since the last restart)
-
-A fifth layer, `tasks`, global only: `global/tasks/` holds one JSON file per DSH scheduled task
-plus a regenerated `index.md`. It backs up the schedule store, which is otherwise the only copy of
-your automated tasks and is in no git repository. The host DELETES a task when its owning session
-stops, so a task can vanish with nothing to observe; the backup runs on the sync cadence and a
-restore re-arms what is missing at the first session after boot.
-
-Verified on disk: 396 tests pass, and `@try-works/dsh-memory@0.1.16` is published and installed
-byte-identical.
-
-- `backupTasks` against `E:\dsh-memory`: two cron tasks captured, a second pass `written=0`.
-- The layer is at `global/tasks/`. It first shipped at `<root>/tasks`; every test used a
-  self-consistent temp root and could not see the difference (`learning-0213`).
-- **The emptied-store drill, on a COPY of the real `schedule.json`:** emptied, backed up, restored
-  through the schedule plugin's own `createCronScheduleRecord`. Both tasks came back with their
-  expressions, zones and prompts (3597 and 2678 chars). The SECOND boot reported `add=0 present=2` —
-  which is the property that matters, since `create` mints a fresh id every time and id-based
-  de-duplication would duplicate every task on every boot.
-- The record -> request -> record round trip holds against the plugin's own builder; only the id
-  differs.
-- No task file leaks into memory search, checked by building the index over a store holding the layer.
-- `SYSTEM.md` names the layer, its path, `memory_tasks`, and all eight states.
-
-**NOT verified, and the log proves it.** `E:\dsh-memory\logs\memory.log` holds 3976 entries and
-ZERO `tasks-backup` or `tasks-restore` ones, while the running host is 3h24m old and predates every
-install from 0.1.10 on. The plugin is alive — it is injecting, capturing and consolidating — it has
-simply never executed this code. The cadence and the boot restore have only ever run inside the
-wiring tests, where the schedule service is a stub and the store is temporary. **Calling the module
-directly is not the same claim as the plugin calling it.**
-
-Seven defects were found by asking what the REAL path does, and not one by an existing test: the
-empty-string config fallback, the wrong directory, a fixture-coincidence assertion, `orphaned` with
-no producer, `skipped` never written while unknown-kind tasks were dropped from the backup, an
-exits table read by nothing and wrong twice, and a manual restore that returned silently while
-reporting success.
-
-Known limitation, by design: the host deletes a task when its owning session stops, so a restored
-task lives as long as the session that adopted it. The boot restore is what re-arms them.
-
-## NOT VERIFIED — treat these as unknown
-
-| item | what is actually known | what would settle it |
-|---|---|---|
-| Session CONTENT search | The index database exists at `~/.dsh/storages/session-search.sqlite` and was opened, so `openAt: first-search` took effect. It holds **0 sessions, 0 documents, `global_generation` 0**. Every sidebar result therefore comes from a browser-side TITLE match. | A **content-only** hit: a query matching text inside a session that appears in no title or workspace name. |
-| Cause of the empty index | Reconciliation runs `_observeStable -> persistence.list -> readColdSessionLog` and indexes nothing. Candidates, untraced: persistence binding unavailable, cold reads failing silently, live sessions skipped. | Tracing the reconcile path with a real store. |
-| Sidebar behaviour with a populated index | The service uses FTS5 MATCH and the client merges content hits with local title matches. The merge code is real; it has never had content to merge. | The same content-only hit above. |
-
-### Where the sidebar search actually comes from
-
-Recorded because it was got wrong once. `client/ui-workspace/src/client/tree.ts`
-`deriveSearchResults` has two sources:
-
-```js
-// LOCAL - browser-side substring match, computed from data already loaded
-sessionTitle(summary).toLowerCase().includes(q)
-  || labelOf(summary).toLowerCase().includes(q)
-
-// CONTENT - server-side FTS hits, empty today
-for (const item of content.items) include(summary)
-```
-
-So title search works and content search is wired but starved. The distinction matters: turning
-the flag on did not change what the user sees.
-
-## TODO — ordered
-
-### Blocked on a restart
-
-- [x] **Restart, then check the tasks layer — DONE 2026-10-09.** The restart confirmed it. See the
-      tasks-layer section above for the two log entries and the counts. It writes to ANOTHER plugin's storage through
-      that plugin's service, so it deserves a real observation rather than a passing suite. After
-      the restart, ask for `memory_tasks` and expect a report with `live: 2`, `backedUp: 2` and
-      `pending.present: 2`. Then check `logs/memory.log` for a `tasks-backup` entry and a
-      `tasks-restore` entry: the first should say `refreshed: 2`, the second `present: 2, added: 0`.
-      An `added` above zero would mean de-duplication failed and it duplicated a task.
-- [ ] **Restart again.** The previous restart worked: `memory_recover` went from "unavailable in
-      this context" to executing. It now fails one layer further on, at output validation, which
-      is fixed on disk. Confirm by calling `memory_recover({})` and getting a report, not an error.
-
-### Recovery (needs the restart)
-
-- [ ] `memory_recover({})` — report only, no cost.
-- [ ] `memory_recover({ limit: 1 })` — **prove it end to end on one** before spending on 76.
-- [ ] Then the rest in batches of ~10. Cost is roughly 76-150 model calls plus consolidation,
-      against **52 MB** of logs. All 76 are under the 8 MB approval threshold, so nothing will
-      stop an accidental `limit: 76` — the one-candidate proof is the guard, not the threshold.
-### Open work, not blocked
-
-- [ ] **Held backfills: 28.** `state/backfill-hold.json`. Released one as a gate test and it
-      produced 2 entries against a 1-3 target, where the live rate was 7-10 — so the conservative
-      bar works. Release in batches and watch the rate after each; ~7 per digest means the bar is
-      not reaching the prompt and the batch should stop.
-- [ ] **79 held records, reason `unclassified`.** Legacy records whose summary was empty and
-      whose cause was never recorded, so a conclusion and a defect are indistinguishable. Parked
-      deliberately: retrying them unasked would be a backfill nobody requested. An operator
-      decision, not a bug.
-- [ ] **Session content search.** See NOT VERIFIED. Two routes: trace why reconciliation indexes
-      nothing, or accept title search and close it. Do not claim it works until a content-only
-      hit is observed.
-- [ ] **Agent-facing session tools** (`session_search` and friends). NOT a config change. In a
-      Web profile the tool plane belongs to a preset, and a patch REPLACES a preset row config
-      wholesale — so adding one tool means restating the whole ~20-row `config.plugins` list and
-      re-syncing by hand whenever the shipped preset changes. Decide this only if the capability
-      is actually missed.
-- [ ] **Pre-`inject`-fix metric distortion is cleared**; `legacy.inject` records the old value
-      and what it meant. Nothing to do, noted so the number is not misread again.
-
-## Known bugs, fixed, and how each was found
-
-Kept because the pattern matters more than the list: every one of these was found by asking
-"does this actually work?" rather than by a test going red.
-
-| bug | how it surfaced |
-|---|---|
-| Retry loop starved | reading the live due list: the retry sat at position 101 of 111 |
-| `memory_recover` unwired | calling the tool; it threw "unavailable in this context" |
-| Wiring test could not catch it | it copied the extras list AND only checked registration |
-| Session content search believed working | reading the index with `node:sqlite`; 0 rows |
-| `inject` metric counted characters | 218,590 "injections" in a store with 101 |
-| 78 orphans misclassified as parked | the migration checked the digest only when unsettled |
-| `orphaned` treated as retryable | it would have burned 78 model calls on missing material |
-| A gate that proved nothing | it asserted guidance present only when a backfill is in the batch |
-| `memory_recover` returned invalid JSON | the harness rejected it; the tests called `execute()` directly and never checked the shape |
-| A test edit silently discarded | my own script re-read and overwrote the file before writing |
-## Published
-
-**`@try-works/dsh-memory@0.1.16`** on npm (0.1.0 was the first release; the Memory tab and the
-tasks layer each added several). The unscoped name `dsh-memory` was already taken by an
-unrelated DeepSeek Harness memory plugin, so the package publishes under the existing `@try-works`
-scope.
-
-What did NOT change, deliberately: the Cordis plugin name (`export const name = 'dsh-memory'`) and
-the bundle row id (`id: dsh-memory`). Config is addressed by ROW ID, so an existing profile's
-`memoryRoot` row survives the rename untouched. Only the package name and the module specifier in
-`cordis.patch.yml` moved.
-
-Verified after publishing: a clean install in a temp directory resolves 0.1.0, ships 49 files and
-42 src modules with no test directory, and the installed entry exports `name`, `inject`, `apply`,
-`Config` and `hostExtras`.
-
-Note for a first publish of any new name: npm creates a `0.0.0-stage` placeholder while the real
-version is staged, so an immediate registry check 404s. It took about three minutes to appear.
-
-## Layout
-
-```
-src/
-  index.js        plugin entry: Config schema, apply(), hostExtras(host)
-  host.js         hooks: session/event, agent/pre-step, session/disposed, idle ticker
-  queue.js        THE state store: states, transitions, retry, watchdog, migration
-  capture.js      session -> digest -> summary; reads and writes queue states
-  summarize.js    the summary call and its reply parsing
-  digest.js       projection: strips injected context, extracts candidates
-  consolidate.js  prompt construction, the line protocol, backfill guidance
-  consolidate-run.js  the consolidation pass
-  recover.js      orphan recovery: plan (dry run) and rebuild-one
-  entry.js        entry formatting, id allocation, parsing
-  layers.js       layered files, topic files, indexes
-  remember.js     the canonical write path
-  forget.js       tombstones
-  refresh.js      freshness re-examination
-  inject.js       the per-turn context card
-  select.js       relevance selection with a recency fallback
-  fts.js          FTS5 index: word + trigram
-  search.js       query -> ranked hits with citations
-  rank.js         rank blending
-  ledger.js       the derived entry ledger
-  state.js        state.md: replace-per-section, freshness
-  working.js      working.md: the agent file
-  tools.js        the memory_* tool surface
-  telemetry.js    metrics and the run log
-  defaults.js     config defaults and validation
-  models.js       the model ladder
-  client/         the Settings page (browser half)
-```
-
-Conventions: plain ESM, no build step, `node --test`. The only dependency is `node:sqlite`,
-which is built into Node.
-
-## Documents
-
-| file | what it is |
-|---|---|
-| **STATUS.md** | this file: what works, what does not, what is left |
-| `README.md` | the plugin as a product: install, configure, tools |
-| `PLAN-queue-wiring.md` | the completed nine-phase plan that made the queue the state store |
-| `dsh-memory-v2-plan.md` | the layered-store design |
-| `old-memory-plan.md` | the retired prototype. Superseded; kept for archaeology only |
-| `BRIEF-sandbox-acl-fix.md` | an unrelated Windows sandbox note |
-
-## The discipline this project settled on
-
-Written down because it is the thing that actually caught the bugs:
-
-1. **Reintroduce the bug to prove the guard.** A test that has never failed has not been shown
-   to test anything. Every guard added late in this project was verified by breaking the thing
-   it protects.
-2. **Verify at the level the user experiences it.** A config flag taking effect is not the
-   feature working. The database existed and the index was empty.
-3. **Prefer a state to a flag.** A reader should never have to derive what happened from a
-   timestamp plus a counter.
-4. **Fail closed on cost.** An unknown size asks for approval; `unavailable` from the approval
-   service means no.
-5. **A bounded consumer turns ordering into policy.** Whatever sorts last may never run.
-6. **Say what is measured and label what is inferred.** The difference is in the tables above.
-7. **A changed error is evidence of progress.** "Unavailable" becoming "invalid output" means the
-   first fix landed and the failure moved one layer deeper. Reading the message rather than the
-   success flag is what located three of the bugs above.
+# dsh-memory — STATUS
+
+A tracker and planner: what exists, what is verified working, what is written but not
+loaded, and what is left. Kept as a record of MEASURED state, not intention.
+
+**Last measured: 2026-10-06** (347 tests passing, 41 src files, 50 test files).
+
+## How to read this
+
+Four states, and the difference between them is the whole point of the document:
+
+| mark | meaning |
+|---|---|
+| **LIVE** | verified working in the running host, by observing an effect it produced |
+| **WRITTEN** | implemented and tested on disk; the running host has not loaded it |
+| **NOT VERIFIED** | believed to work; no observation confirms it. Treat as unknown |
+| **TODO** | not done |
+
+This distinction exists because of two failures in this project. A retry loop was reported
+done and was logically correct but could never run, starved behind a backlog in a bounded
+batch. And session content search was reported enabled on the strength of a config flag; the
+database was created and the index was empty. In both cases a green result masked a dead
+feature. Both are recorded below as they actually are.
+
+## What dsh-memory is
+
+A DSH plugin storing durable agent memory as layered Markdown under a configurable root, with
+a global layer and per-workspace layers, committable to a git origin. Markdown is the record;
+indexes and state are derived and gitignored.
+
+| layer | scope | contents |
+|---|---|---|
+| `semantic` | global + workspace | facts, preferences, standing rules |
+| `episodic` | global + workspace | what happened, session by session |
+| `procedural` | global + workspace | how to do things here |
+| `learning` | global + workspace | lessons, usually from bugs |
+| `decision` | global + workspace | settled choices with their reasons |
+| `codebase` | workspace only | what each module does |
+| `history` | workspace only | commits, deduped by content |
+| `state.md` | workspace only | the repository now, replaced per section, freshness-marked |
+| `working.md` | workspace only | the agent now: task, activity, next steps |
+
+Live store: **1200 entries** — decision 264, learning 280, semantic 278, codebase 173,
+procedural 103, episodic 62, history 40.
+## Built and LIVE
+
+Observed in the running host. Each row names the effect that was seen, not a claim.
+
+| capability | evidence |
+|---|---|
+| Queue is the only state store | `state/retired/sessions.json`; `state/queue.jsonl` holds 170 entries |
+| Capture writes transitions | queue states move with every session; `summarized` 19 → 23 since boot |
+| Consolidation writes transitions | `consolidate` events fire; digests reach `added` |
+| Retry scheduling | one entry is `failed` with a reason; `due()` returns it as `retry` |
+| Stalled-stage watchdog | `reapStalled` runs before each batch; 0 stranded mid-stage |
+| Explicit abandonment | states drain rather than accumulate; `ABANDONED` reported by `problems()` |
+| `memory_status` reports the queue | per-state counts, parked reasons, grouped attention with ages |
+| Attention is GROUPED | 2 rows where the first version produced 79, 78 of them identical |
+| Injection selects by RELEVANCE | the context card reads "Relevant learnings" and lists this session's work, where it previously always listed UI notes |
+| History ingest | `history-check` events, 5 since the instrumented host started |
+| Injection metric counts injections | `events.inject` advances by 1 per injection; the old character total is preserved under `legacy.inject` |
+| `memory_recover` is wired | calling it no longer answers "unavailable in this context"; it reaches output validation |
+| Session logs are searchable in principle | zstd, not encrypted; 368 logs, 740MB compressed, ~2.96GB plaintext |
+
+## WRITTEN but NOT LOADED
+
+Implemented and tested on disk. The running host predates them, so none of this is in effect.
+
+| change | why it matters | how it was verified on disk |
+|---|---|---|
+| `due()` priority ordering | retry → abandon → work → recover, so a bounded batch reaches a due retry | measured: the retry sat at position 101 of 111 and is now position 1 |
+| `hostExtras(host)` as one source | `index.js` and the wiring test shared a copied list that drifted | the test now derives extras from the host object |
+| `memory_recover` report omits its `entries` field rather than blanking it | the old line added `entries: undefined`, which is NOT valid JSON, so the harness rejected the whole result | the new guard round-trips every tool output; reintroducing it fails with `value.entries is undefined` |
+
+How to tell loaded from not: ask the host. `memory_recover` used to reply *"unavailable in this
+context"* and now replies *"returned invalid output: value is not lossless JSON"*. The change of
+ERROR is the proof the first fix is loaded and the tool now executes — a different failure, one
+layer further along. The report fix in the last row is what remains pending.
+### The tasks layer (new since the last restart)
+
+A fifth layer, `tasks`, global only: `global/tasks/` holds one JSON file per DSH scheduled task
+plus a regenerated `index.md`. It backs up the schedule store, which is otherwise the only copy of
+your automated tasks and is in no git repository. The host DELETES a task when its owning session
+stops, so a task can vanish with nothing to observe; the backup runs on the sync cadence and a
+restore re-arms what is missing at the first session after boot.
+
+Verified on disk: 396 tests pass, and `@try-works/dsh-memory@0.1.16` is published and installed
+byte-identical.
+
+- `backupTasks` against `E:\dsh-memory`: two cron tasks captured, a second pass `written=0`.
+- The layer is at `global/tasks/`. It first shipped at `<root>/tasks`; every test used a
+  self-consistent temp root and could not see the difference (`learning-0213`).
+- **The emptied-store drill, on a COPY of the real `schedule.json`:** emptied, backed up, restored
+  through the schedule plugin's own `createCronScheduleRecord`. Both tasks came back with their
+  expressions, zones and prompts (3597 and 2678 chars). The SECOND boot reported `add=0 present=2` —
+  which is the property that matters, since `create` mints a fresh id every time and id-based
+  de-duplication would duplicate every task on every boot.
+- The record -> request -> record round trip holds against the plugin's own builder; only the id
+  differs.
+- No task file leaks into memory search, checked by building the index over a store holding the layer.
+- `SYSTEM.md` names the layer, its path, `memory_tasks`, and all eight states.
+
+**NOT verified, and the log proves it.** `E:\dsh-memory\logs\memory.log` holds 3976 entries and
+ZERO `tasks-backup` or `tasks-restore` ones, while the running host is 3h24m old and predates every
+install from 0.1.10 on. The plugin is alive — it is injecting, capturing and consolidating — it has
+simply never executed this code. The cadence and the boot restore have only ever run inside the
+wiring tests, where the schedule service is a stub and the store is temporary. **Calling the module
+directly is not the same claim as the plugin calling it.**
+
+Seven defects were found by asking what the REAL path does, and not one by an existing test: the
+empty-string config fallback, the wrong directory, a fixture-coincidence assertion, `orphaned` with
+no producer, `skipped` never written while unknown-kind tasks were dropped from the backup, an
+exits table read by nothing and wrong twice, and a manual restore that returned silently while
+reporting success.
+
+Known limitation, by design: the host deletes a task when its owning session stops, so a restored
+task lives as long as the session that adopted it. The boot restore is what re-arms them.
+
+## NOT VERIFIED — treat these as unknown
+
+| item | what is actually known | what would settle it |
+|---|---|---|
+| Session CONTENT search | The index database exists at `~/.dsh/storages/session-search.sqlite` and was opened, so `openAt: first-search` took effect. It holds **0 sessions, 0 documents, `global_generation` 0**. Every sidebar result therefore comes from a browser-side TITLE match. | A **content-only** hit: a query matching text inside a session that appears in no title or workspace name. |
+| Cause of the empty index | Reconciliation runs `_observeStable -> persistence.list -> readColdSessionLog` and indexes nothing. Candidates, untraced: persistence binding unavailable, cold reads failing silently, live sessions skipped. | Tracing the reconcile path with a real store. |
+| Sidebar behaviour with a populated index | The service uses FTS5 MATCH and the client merges content hits with local title matches. The merge code is real; it has never had content to merge. | The same content-only hit above. |
+
+### Where the sidebar search actually comes from
+
+Recorded because it was got wrong once. `client/ui-workspace/src/client/tree.ts`
+`deriveSearchResults` has two sources:
+
+```js
+// LOCAL - browser-side substring match, computed from data already loaded
+sessionTitle(summary).toLowerCase().includes(q)
+  || labelOf(summary).toLowerCase().includes(q)
+
+// CONTENT - server-side FTS hits, empty today
+for (const item of content.items) include(summary)
+```
+
+So title search works and content search is wired but starved. The distinction matters: turning
+the flag on did not change what the user sees.
+
+## TODO — ordered
+
+### Blocked on a restart
+
+- [x] **Restart, then check the tasks layer — DONE 2026-10-09.** The restart confirmed it. See the
+      tasks-layer section above for the two log entries and the counts. It writes to ANOTHER plugin's storage through
+      that plugin's service, so it deserves a real observation rather than a passing suite. After
+      the restart, ask for `memory_tasks` and expect a report with `live: 2`, `backedUp: 2` and
+      `pending.present: 2`. Then check `logs/memory.log` for a `tasks-backup` entry and a
+      `tasks-restore` entry: the first should say `refreshed: 2`, the second `present: 2, added: 0`.
+      An `added` above zero would mean de-duplication failed and it duplicated a task.
+- [ ] **Restart again.** The previous restart worked: `memory_recover` went from "unavailable in
+      this context" to executing. It now fails one layer further on, at output validation, which
+      is fixed on disk. Confirm by calling `memory_recover({})` and getting a report, not an error.
+
+### Recovery (needs the restart)
+
+- [ ] `memory_recover({})` — report only, no cost.
+- [ ] `memory_recover({ limit: 1 })` — **prove it end to end on one** before spending on 76.
+- [ ] Then the rest in batches of ~10. Cost is roughly 76-150 model calls plus consolidation,
+      against **52 MB** of logs. All 76 are under the 8 MB approval threshold, so nothing will
+      stop an accidental `limit: 76` — the one-candidate proof is the guard, not the threshold.
+### Open work, not blocked
+
+- [ ] **Held backfills: 28.** `state/backfill-hold.json`. Released one as a gate test and it
+      produced 2 entries against a 1-3 target, where the live rate was 7-10 — so the conservative
+      bar works. Release in batches and watch the rate after each; ~7 per digest means the bar is
+      not reaching the prompt and the batch should stop.
+- [ ] **79 held records, reason `unclassified`.** Legacy records whose summary was empty and
+      whose cause was never recorded, so a conclusion and a defect are indistinguishable. Parked
+      deliberately: retrying them unasked would be a backfill nobody requested. An operator
+      decision, not a bug.
+- [ ] **Session content search.** See NOT VERIFIED. Two routes: trace why reconciliation indexes
+      nothing, or accept title search and close it. Do not claim it works until a content-only
+      hit is observed.
+- [ ] **Agent-facing session tools** (`session_search` and friends). NOT a config change. In a
+      Web profile the tool plane belongs to a preset, and a patch REPLACES a preset row config
+      wholesale — so adding one tool means restating the whole ~20-row `config.plugins` list and
+      re-syncing by hand whenever the shipped preset changes. Decide this only if the capability
+      is actually missed.
+- [ ] **Pre-`inject`-fix metric distortion is cleared**; `legacy.inject` records the old value
+      and what it meant. Nothing to do, noted so the number is not misread again.
+
+## Known bugs, fixed, and how each was found
+
+Kept because the pattern matters more than the list: every one of these was found by asking
+"does this actually work?" rather than by a test going red.
+
+| bug | how it surfaced |
+|---|---|
+| Retry loop starved | reading the live due list: the retry sat at position 101 of 111 |
+| `memory_recover` unwired | calling the tool; it threw "unavailable in this context" |
+| Wiring test could not catch it | it copied the extras list AND only checked registration |
+| Session content search believed working | reading the index with `node:sqlite`; 0 rows |
+| `inject` metric counted characters | 218,590 "injections" in a store with 101 |
+| 78 orphans misclassified as parked | the migration checked the digest only when unsettled |
+| `orphaned` treated as retryable | it would have burned 78 model calls on missing material |
+| A gate that proved nothing | it asserted guidance present only when a backfill is in the batch |
+| `memory_recover` returned invalid JSON | the harness rejected it; the tests called `execute()` directly and never checked the shape |
+| A test edit silently discarded | my own script re-read and overwrote the file before writing |
+## Published
+
+**`@try-works/dsh-memory@0.1.16`** on npm (0.1.0 was the first release; the Memory tab and the
+tasks layer each added several). The unscoped name `dsh-memory` was already taken by an
+unrelated DeepSeek Harness memory plugin, so the package publishes under the existing `@try-works`
+scope.
+
+What did NOT change, deliberately: the Cordis plugin name (`export const name = 'dsh-memory'`) and
+the bundle row id (`id: dsh-memory`). Config is addressed by ROW ID, so an existing profile's
+`memoryRoot` row survives the rename untouched. Only the package name and the module specifier in
+`cordis.patch.yml` moved.
+
+Verified after publishing: a clean install in a temp directory resolves 0.1.0, ships 49 files and
+42 src modules with no test directory, and the installed entry exports `name`, `inject`, `apply`,
+`Config` and `hostExtras`.
+
+Note for a first publish of any new name: npm creates a `0.0.0-stage` placeholder while the real
+version is staged, so an immediate registry check 404s. It took about three minutes to appear.
+
+## Layout
+
+```
+src/
+  index.js        plugin entry: Config schema, apply(), hostExtras(host)
+  host.js         hooks: session/event, agent/pre-step, session/disposed, idle ticker
+  queue.js        THE state store: states, transitions, retry, watchdog, migration
+  capture.js      session -> digest -> summary; reads and writes queue states
+  summarize.js    the summary call and its reply parsing
+  digest.js       projection: strips injected context, extracts candidates
+  consolidate.js  prompt construction, the line protocol, backfill guidance
+  consolidate-run.js  the consolidation pass
+  recover.js      orphan recovery: plan (dry run) and rebuild-one
+  entry.js        entry formatting, id allocation, parsing
+  layers.js       layered files, topic files, indexes
+  remember.js     the canonical write path
+  forget.js       tombstones
+  refresh.js      freshness re-examination
+  inject.js       the per-turn context card
+  select.js       relevance selection with a recency fallback
+  fts.js          FTS5 index: word + trigram
+  search.js       query -> ranked hits with citations
+  rank.js         rank blending
+  ledger.js       the derived entry ledger
+  state.js        state.md: replace-per-section, freshness
+  working.js      working.md: the agent file
+  tools.js        the memory_* tool surface
+  telemetry.js    metrics and the run log
+  defaults.js     config defaults and validation
+  models.js       the model ladder
+  client/         the Settings page (browser half)
+```
+
+Conventions: plain ESM, no build step, `node --test`. The only dependency is `node:sqlite`,
+which is built into Node.
+
+## Documents
+
+| file | what it is |
+|---|---|
+| **STATUS.md** | this file: what works, what does not, what is left |
+| `README.md` | the plugin as a product: install, configure, tools |
+| `PLAN-queue-wiring.md` | the completed nine-phase plan that made the queue the state store |
+| `dsh-memory-v2-plan.md` | the layered-store design |
+| `old-memory-plan.md` | the retired prototype. Superseded; kept for archaeology only |
+| `BRIEF-sandbox-acl-fix.md` | an unrelated Windows sandbox note |
+
+## The discipline this project settled on
+
+Written down because it is the thing that actually caught the bugs:
+
+1. **Reintroduce the bug to prove the guard.** A test that has never failed has not been shown
+   to test anything. Every guard added late in this project was verified by breaking the thing
+   it protects.
+2. **Verify at the level the user experiences it.** A config flag taking effect is not the
+   feature working. The database existed and the index was empty.
+3. **Prefer a state to a flag.** A reader should never have to derive what happened from a
+   timestamp plus a counter.
+4. **Fail closed on cost.** An unknown size asks for approval; `unavailable` from the approval
+   service means no.
+5. **A bounded consumer turns ordering into policy.** Whatever sorts last may never run.
+6. **Say what is measured and label what is inferred.** The difference is in the tables above.
+7. **A changed error is evidence of progress.** "Unavailable" becoming "invalid output" means the
+   first fix landed and the failure moved one layer deeper. Reading the message rather than the
+   success flag is what located three of the bugs above.