# Changelog

All notable changes to this project are documented here.

## [0.10.0] - 2026-10-10

### DSH desktop 0.2

- Export a volatile `Config` so desktop `SettingsForms` projects
  `memory.enabled`; read it from the loader-owned config, apply edits live via
  `loader/volatile-update`, and write through `SettingsForms.update(<entry id>)`.
  The legacy `settings.register()` scope is kept for the 0.1.x web runtime.
- Bind the optional settings service whenever it becomes available instead of
  sampling it once at startup; a write without any settings backend now fails
  with `settings-unavailable` instead of reporting success, and an unreadable
  setting fails closed.
- Widen peers to `>=0.2.0-rc.2 <0.3.0`; the aggregate bundle declares the
  settings-slots client seam like the standalone UI package.
- The synchronizer honors `memory.enabled` from the desktop profile patch and
  falls back to the desktop app's bundled CLI when `dsh` is not on `PATH`.

### Sync

- Discover every session-log generation (`session.v4.jsonl.zstd` on desktop)
  and deliver only the newest per session; nothing was synchronized from
  desktop sessions before this fix.
- Skip live sessions whose `session.lock` is held, the synchronizer's own
  staging sessions, and tmp working directories (`DSH_MEMORY_SYNC_INCLUDE_TMP=1`
  opts tmp back in).
- Generalize the pending-generation fence (exit 78) to any older generation.
- Write a metadata-only `.sync/heartbeat.json` on every real run.
- Journal commits stage only `.sync/runs/<id>.json`, `.sync/last-run.json` and
  `.last-sync`, never runtime state such as the operation lock.

### Fixed

- `clearActiveRun` compared camelCase `runId` against the snake_case `run_id`
  the host writes, stranding `active-run.json` after UI migrate/rollback/clear
  and blocking every later sync with `operation-in-progress`.
- Hybrid search ranked the raw lexical component by file-name order; it now
  ranks by lexical score.
- `memory_context` with a query keeps relevance order instead of re-sorting by
  usage; the default active scope excludes superseded/archived records without
  a query too.
- Rollback refuses a dirty payload worktree (`rollback-dirty-worktree`) and
  restores HEAD, index and payload if it fails after moving HEAD.
- The memory prompt section is not template-expanded (`interpolate: false`,
  braces neutralized), refreshes are generation-ordered (no duplicate or
  resurrected section), and the absolute memory path is no longer in the prompt.
  An empty `summary.md` yields a short notice instead of an empty snapshot.
- `memory_search`/`memory_context` return `memory-disabled` while memory is off,
  bound `limit` to 1–20 with an integer schema, constrain `scope` to an enum,
  and omit ranking internals and hashes from model-facing output.

### Tests and tooling

- `run-memory-tests.sh` runs every suite and reports all failures; the runtime
  test checks module behavior instead of exact source text.
- New: host contract, ranking, generation discovery, desktop CLI resolution.
- `test:desktop-compat` loads the plugins against a real desktop dependency
  graph (`DSH_DESKTOP_NODE_MODULES`); CI runs runtime contracts for 0.1.5-rc.1
  and 0.2.0-rc.2, with cordis/schemastery taken from that release's metadata.
- `npm test` fails on marketplace bundle drift.

### Earlier unreleased compatibility work

- Verify Host and client entry points against DSH desktop 0.1.7-rc.1; remove the legacy client-runtime injection from the memory UI.
- Adapt Host entry points and peer ranges to DSH 0.1.5-rc.1.
- Add real-runtime import coverage while retaining legacy tests.
- Scan V3 sessions without duplicate V2 processing; fence pending generation changes.


## [0.9.3] - 2026-09-20

### Fixed

- Retrieve unsegmented Chinese queries. Every CJK character matches `\p{L}`, so
  the query tokenizer collapsed a whole Chinese phrase into a single token;
  `rawLexicalScore` matches by substring and `retrieve` discards records scoring
  zero, so a Chinese query without spaces returned nothing at all (measured:
  `记忆系统的已知缺陷与优化路线` returned 0 hits while `记忆 系统 缺陷` returned 74).
  Query tokens now also expand into overlapping CJK bigrams, so long Chinese
  queries match by adjacent-pair overlap and rank by hit count. Non-CJK queries
  are unchanged.
- Accept unindented block lists in record front matter. The staging validator
  (`sync-apply.py`) and the plugin validator (`memory-metadata.js`) both required
  indentation on list items and rejected valid YAML with `invalid-metadata`,
  although the metadata contract only forbids flow-style collections. Both now
  accept either block form while still rejecting `[a, b]` and block scalars.
- Name the offending path, field and line when front matter is rejected, so a
  failing run identifies the file instead of reporting a bare error code.
- Raise `maxBuffer` for the Git index snapshot in the public-tree guard.
  `execFileSync` defaults to 1 MiB, so the guard failed with
  `cannot construct Git index snapshot` once the tree grew past that size.
- Probe the Playwright browser binary before the E2E acceptance test. The guard
  only checked that `python3` exists, so a Playwright upgrade without a matching
  `playwright install` broke `npm test` instead of skipping as documented.

### Added

- `integrations/dsh/dsh-memory-housekeeping.py`: a deterministic, idempotent
  host-side pass that adds minimal schema v1 front matter to legacy records
  (mirroring the plugin's `legacy-migration.js` id derivation) and archives
  superseded or long-unreferenced records **in place** through
  `status: archived`. Records are never moved: `source_rollouts` must start with
  `rollouts/`, so relocating one into `archive/` would dangle every reference to
  it.

### Changed

- `integrations/dsh/dsh-memory-sync` now states the permitted front matter values
  (`type`/`status`/`confidence` enumerations) and a `summary.md` writing budget
  with a 6 KiB soft target plus a one-line-per-project rule for the
  recent-terminal-state section, and drops only the summary change when it
  exceeds the hard budget instead of rejecting the whole diff. The operation lock
  is keyed to the sync process rather than a short-lived helper, so an
  interrupted run no longer blocks the next one for six hours.

## [Unreleased]

## [0.9.2] - 2026-09-01

### Fixed

- Quote FTS5 search terms before constructing `MATCH` expressions so
  hyphenated identifiers (for example, `pending-candidates`) stay on the
  hybrid index path instead of being parsed as SQLite expressions.
- Add a regression case to the retrieval evaluation for hyphenated queries.

### Distribution

- Distributed through GitHub source installs and GitHub Releases; not
  published to npm.

## [0.9.1] - 2026-09-01

### Added

- Explicit-only conflict resolution: `resolveTopicConflict()` no longer
  applies implicit "newest wins" ordering. A winner emerges only through an
  explicit `supersedes` link (the named record is projected out), records
  linked by `conflicts_with` block each other until an operator resolves the
  conflict, and zero-or-multiple contenders resolve to "unresolved" instead
  of silently discarding knowledge. Expired records and records whose own
  status is `superseded`/`archived` still project out. Mirrored in the host
  staging helper.
- Handbook provenance gate (fail-closed): newly added `handbook/` records in
  a staged diff must carry at least one of `source_rollouts`,
  `source_session_digest`, or `source_hash`; the host rejects the whole diff
  with `missing-provenance` otherwise. Baseline (pre-existing) records are
  exempt so legacy migration keeps working. The consolidation prompt and the
  managed template README document the requirement and the explicit
  supersession/conflict workflow.
- Independent enrichment preview: `--preview` metadata now splits
  `changed_paths` into `enrichment_paths` (under `handbook/`) and
  `consolidation_paths`, and the run output reports both categories so
  distilled handbook knowledge can be reviewed separately from
  consolidation edits. `--apply-preview` still applies the captured diff
  atomically. Staging-diff failures now print structured error details
  (path/phase/run id) for faster diagnosis.
- Release acceptance checklist: `docs/release-checklist.md` gains the local
  deployment acceptance sequence — backup, quiesce the memory-sync
  LaunchAgent only, safe-window DSH restart, zero-Provider `--scan-only
  --json` acceptance, scheduler reload, and a hybrid-retrieval smoke check.

### Changed

- Default retrieval scope: the `active` scope (and the legacy read path)
  now excludes records whose own status is `superseded` or `archived`;
  expired records were already skipped. Retired knowledge remains
  retrievable through `scope: "all"` or the archive directory.
- `SYNC_POLICY_VERSION` advances to `v0.9.1`: the staging contract changed
  (provenance gate, explicit conflict chain).

## [0.9.0] - 2026-09-01

### Added

- Hybrid retrieval pipeline: `memory.search()` and `memory.context()` now fuse
  a full-corpus lexical scan with a derived SQLite full-text index. Raw
  lexical rank, FTS rank, coverage rank, and usage rank are combined with
  reciprocal-rank fusion (k=60, weights raw 2.0 / FTS 1.0 / coverage 1.0 /
  usage 0.25), and exact-substring matches keep a dedicated precedence tier
  above all fused scoring.
- Derived search index at `.sync/search-index.sqlite` (permission `0600`):
  a `content_fts` table with the trigram tokenizer (CJK runs of 3+ characters
  and latin words of 3+ characters) plus a `terms_fts` table with `unicode61`
  terms over English words and CJK bigrams. The index is stamped with
  `sha256(payloadTree(HEAD))`, built into a temporary database that is atomically
  renamed into place, and kept out of the working tree via `.git/info/exclude`.
- Retrieval API metadata: responses gain `retrieval`
  (`mode`: `hybrid` or `scan`; `indexState`: `fresh`, `rebuilt`, or
  `degraded`; `stamp`), and results gain `score` plus `score_components`
  (`raw`, `fts`, `coverage`, `usage` contributions) in both search and
  query-aware context responses.
- Transparent degradation: when the index is missing, stale, or corrupt it is
  rebuilt automatically; if rebuilding fails (for example an unwritable
  `.sync` directory) retrieval falls back to the pure scan with
  `indexState: "degraded"` and identical results.
- Retrieval evaluation suite (`tests/test_dsh_memory_retrieval_eval.mjs`):
  52 labeled queries over a 26-record mixed Chinese/English corpus asserting
  top-1/top-3 recall, scope filtering, exact-tier precedence over usage
  reinforcement, usage-unchanged invariants, index permissions, working-tree
  cleanliness,   stamp invalidation on new commits, and corrupt-index
  degradation.

## [0.8.4] - 2026-09-01

### Added

- Raw caps and model input caps are now separate budgets: raw session files
  (`DSH_MEMORY_MAX_SESSION_FILE_BYTES`, 64 MiB default), decompressed
  payloads (`DSH_MEMORY_MAX_SESSION_DECOMPRESSED_BYTES`, 128 MiB default),
  and per-chunk bytes (`DSH_MEMORY_MAX_CHUNK_BYTES`, 1 MiB default) bound the
  host side, while `DSH_MEMORY_MAX_CANDIDATE_BYTES` (8 MiB default) bounds
  the model input per run.
- Streaming sanitizer: `filter_session.py` decodes zstd/plain session logs
  line by line without materializing the payload, fails closed with
  `decompressed-too-large`, and writes fixed-size sanitized chunks prefixed
  with a provenance header (`session_id`, `session_digest`,
  `chunk_index/chunk_total`).
- Per-candidate delivery with resume: size-class rejections are counted per
  candidate (`rejected_candidate_count`) instead of aborting the batch;
  partially delivered sessions are deferred (`deferred_candidate_count`) and
  recorded in pending-candidates v2 (`digest`, `next_chunk`, `chunk_total`,
  `complete`), so the next run resumes from the first undelivered chunk
  without re-paying the provider for delivered chunks.
- Journal records and scan-only output gain `processed_chunk_count`,
  `deferred_candidate_count`, and `rejected_candidate_count`.

### Changed

- Size-class rejections (`raw cap`, `decompressed-too-large`) no longer fail
  the batch: the watermark is held past unresolved candidates and the run is
  journaled zero-cost (`rejected`/`deferred`) when nothing reaches the model.
  Infrastructure failures (missing decoder, crashed filter, unreadable chunk
  metadata) keep failing the whole batch closed.
- The watermark only advances past a fully resolved window: truncation,
  deferrals, and rejections all hold it in place.
- The consolidation prompt documents the chunked contract (group by
  `session_id`, read chunks in `chunk_index` order, undelivered chunks
  arrive on the next run).

## [0.8.3] - 2026-09-01

### Added

- Metadata audit: `status`/`health` now report `metadataValid`,
  `validMetadataCount`, `invalidMetadataCount`, and up to 20 `invalidMetadata`
  entries. Metadata anomalies surface as a dedicated UI warning ("记忆数据格式
  异常") and never fold into `needsManualRecovery`.
- Search and context exclude records with invalid metadata from retrieval and
  return per-file `warnings` with stable error codes instead of failing.
- Flow-style front matter collections (`tags: [a, b]`) are rejected with the
  stable `flow-style-metadata` code; new/updated records must use canonical
  block lists.
- Single-writer coordination: the memory section now instructs normal sessions
  to treat the memory repository as read-only and record memory requests in
  their session output; the periodic transactional sync performs extraction,
  consolidation, and git commits. README templates document the write boundary
  and the canonical front matter format.
- README.md, `.sync/.gitignore`, and `scripts/filter_session.py` are managed
  readonly templates: identical checksum is a no-op, a dirty live copy fails
  closed, changed files are replaced atomically and committed once as
  "Refresh managed memory templates".
- `--scan-only [--json]` is the zero-Provider acceptance path: it reports
  candidates, raw bytes, digests, pending state, and the watermark without
  starting DSH, acquiring the operation lock, writing a journal, or advancing
  the watermark. `--dry-run` is a deprecated alias that no longer runs the
  model.
- Prompt freshness: the memory section re-renders when `summary.md` changes
  (polled every 5s, debounced, watcher stopped on cleanup, brief file absence
  keeps the previous snapshot).

### Changed

- The memory skill (v1.2.0) turns extract/consolidate/forget into memory
  requests recorded in session output; direct file writes and git commits are
  removed from normal sessions. Forget defaults to archive/correction with
  explicit `supersedes`.

## [0.8.2] - 2026-08-25

### Fixed

- Session filtering now distinguishes zstd frames from plain JSONL by magic
  bytes, works under launchd's minimal `PATH`, and fails closed with
  `zstd-unavailable` when a decoder is missing.
- Existing memory repositories refresh the deployed session helper
  idempotently, refuse dirty helper files, and record helper-only refresh
  commits. This also deploys the template's credential redaction to stale
  local installations.

### Operational notes

- A single session larger than `MAX_CANDIDATE_BYTES` still rejects the whole
  batch; per-file skip-and-journal is tracked separately.
- This release is distributed through GitHub source installs and GitHub
  Releases; it is not published to npm.

## [0.8.1] - 2026-08-24

### Added

- Host-side duplicate-id diagnostics now run against both the live baseline and
  final staging tree, including the conflicting id and paths.
- Failed stage-copy, child, diff, and apply phases write metadata-only journals;
  retry backoff, failure sentinels, lock-rejection logs, and batch limits prevent
  silent repeated Provider work.
- Summary budget checks, archive-aware read scopes, cited content hashes, and
  generation-aware usage metadata are now available without new UI.
- The public `dsh-git-memory` bundle is synchronized from the private host/UI
  workspaces for the v0.8.1 release.

### Changed

- `last-run.json` now represents only a real attempt; `skipped-retry` records
  remain queryable in `.sync/runs/` without overwriting the last attempt.
- Existing legacy ids remain stable logical usage ids; explicit future renames
  can use usage aliases without resetting history.

### Fixed

- Summary injection and consolidation rules enforce a 12 KiB budget.

- LaunchAgent sync now discovers a local Node runtime when launchd provides a
  minimal `PATH`, so the pinned `#!/usr/bin/env node` DSH launcher does not
  fail before headless consolidation.
- Workspace and package lock metadata now consistently records v0.8.0.
- Removed the Legacy migration card and controls from the settings UI; the
  host API and `dsh-memory-migrate` CLI remain available.
- Automatic headless consolidation sessions now use a private per-run
  persistence root, so they do not appear in the normal DSH conversation list.

### Changed

- Reworked the memory settings block into native settings-style accordion
  rows: repository and long-term-memory summaries stay compact, while recent
  syncs, pending previews, and destructive actions expand inline.
- Added keyboard/focus-safe inline confirmations, semantic status tones, and
  empty-preview hiding. Browser E2E now links the source workspaces directly
  instead of exercising an installed profile copy.
- The enabled memory prompt now includes a bounded, explicitly untrusted
  `summary.md` snapshot, with a static-instructions fallback when the file is
  unavailable.

### Added

- Registered host-owned read tools `memory_search` and `memory_context`; both
  return JSON with relative-path citations, and `memory_context` records only
  metadata-only usage feedback.
- Zero-change sync runs now write a `no_change` journal and advance the
  watermark, while dry-run and preview modes remain side-effect free.
- Added the public `dsh-git-memory` root bundle: one DSH install now activates
  the Git-backed memory host and the accordion settings client through a
  standard `dsh.bundle` + `dsh.client` manifest. The internal workspace
  packages remain private.

## [0.8.0] - 2026-08-21

### Added

- `memory.context({ query, limit })` returns bounded memory records with stable
  source citations and deterministic usage-aware ordering.
- Metadata-only usage feedback is stored atomically in the private
  `.sync/usage.json` sidecar (`usage_count`, `last_usage`) and never enters
  payload files or journal records.
- Existing memory repositories add the sidecar to local Git excludes without
  modifying tracked files; new repositories include it in `.sync/.gitignore`.

### Safety

- Context reads are host-owned and read-only with a 20-record limit and a
  4 KiB per-record body bound; UTF-8 truncation is boundary-safe.
- Expired records are excluded, citations expose only relative payload paths
  and optional record ids, and no UI or filesystem path selector was added.
- Usage lock ownership and stale temporary-file cleanup prevent concurrent
  readers from losing counts or leaking sidecar artifacts into Git.

## [0.7.0] - 2026-08-21

### Added

- Browser end-to-end tests for the settings UI: the suite boots a throwaway
  DSH web profile on an ephemeral port (reusing the pinned runtime and shared
  plugin store, symlinking the local workspaces, and seeding onboarding
  settings plus an empty fixture memory repository), then drives headless
  Chromium to assert the panel across desktop/light, dark, and narrow
  viewports, plus empty-state hiding of the Legacy and preview sections and
  the enable switch.
- The E2E suite runs as part of `npm test` and skips cleanly when the Python
  Playwright browser-acceptance tooling is unavailable.

## [0.6.0] - 2026-08-20

### Added

- `memory.legacyRecords()` exposes metadata-only legacy-record discovery to
  the host and settings UI.
- `memory.migrateLegacy({ dryRun })` moves the existing CLI migration logic
  into the library and applies front matter through the normal staged,
  host-owned recovery/apply transaction.
- `dsh-memory-migrate --dry-run|--apply` now delegates to the library API.
- Settings UI now shows pending legacy paths and deterministic generated ids,
  and offers an explicit migration action.

### Safety

- Dry-run does not write the live repository, journal, watermark, or Git.
- Migration preserves legacy record bodies and stores metadata only in the
  journal; no transcript, prompt, credential, or memory body is exposed.
- Applying migration is idempotent and uses the existing operation lock.

## [0.5.0] - 2026-08-20

### Added

- `dsh-memory-backup export <bundle>`: self-contained Git bundle of the full
  memory history plus a manifest sidecar (head, commit count, payload file
  list, size). The live store is never modified.
- `dsh-memory-backup import <bundle> [--target <root>]`: restores a bundle
  into a new directory as a complete Git worktree (recovery/apply/journal/
  rollback history preserved). Refuses an existing target; `--dry-run`
  verifies without writing.
- `docs/compatibility.md`: DSH runtime matrix (rc.6 declared peer, rc.7
  verified baseline), operating-system scope (macOS supported, Linux
  expected, Windows unsupported), and integration tool defaults.
- Release checklist now verifies the backup round-trip before release.

### Compatibility

- DSH `0.1.0-rc.6` remains within the declared peer range.
- DSH `0.1.0-rc.7` is the reproducible clean-room and locally integrated
  baseline.

### Safety

- Backup export never touches the live store; import never overwrites an
  existing repository and verifies the bundle before writing.
- The backup manifest contains metadata only (head, counts, file list) —
  no transcripts, prompts, or credentials.

## [0.4.0] - 2026-08-20

### Added

- Namespaced record ids: a single `/` separator (`project/codegen`,
  `user/preferences`); each segment is `[a-z0-9][a-z0-9-]*`. Enforced by both
  the JS metadata parser and the Python staging validator.
- Optional provenance front matter fields: `source_hash`, `created_by`,
  `review_after`, `expires_at`. Absent fields parse as `null` and render
  canonically; invalid values fail closed.
- Lazy expiry projection: a record with `expires_at` in the past is excluded
  from `search()` and from deterministic conflict resolution without
  rewriting its front matter.
- Deterministic conflict resolution: records sharing a topic key
  (`type:namespace`) pick a winner by status precedence (active > candidate >
  conflicted > superseded > archived), then newest `updated_at`, then smallest
  id. Expired records never win.
- `memory.search({ query, limit })` local full-text search over payload
  records with front matter + body scoring and snippets.
- `memory/search` remote descriptor for the settings UI.

### Compatibility

- DSH `0.1.0-rc.6` remains within the declared peer range.
- DSH `0.1.0-rc.7` is the reproducible clean-room and locally integrated
  baseline.

### Safety

- Malformed provenance values and out-of-pattern namespaced ids fail closed in
  both parsers.
- Lazy expiry never mutates record files; a later consolidation decides how to
  archive or remove expired content.

## [0.3.1] - 2026-08-20

### Added

- `dsh-memory-sync --preview <id>` captures a candidate diff as a pending
  preview under `<root>/.sync/previews/<id>` (7-day expiry) without applying.
- `dsh-memory-sync --apply-preview <id>` applies a pending preview as a normal
  sync transaction and journals it under `operation: preview`; the preview is
  consumed on success.
- `dsh-memory-sync --discard-preview <id>` removes a pending preview.
- `dsh-memory-sync --dry-run --json` emits a single machine-parseable JSON
  report with all progress lines suppressed from stdout.
- `memory.previews()`, `memory.applyPreview()`, and `memory.discardPreview()`
  host APIs, plus settings-UI preview list with apply/discard actions.
- `prepare-preview` now seeds the preview staging from the live payload tree
  (baseline + directory skeleton + manifest), so an applied preview is a
  complete transaction.
- Helper error envelopes are recovered from non-zero helper exits (stdout JSON)
  instead of collapsing to a generic failure code.

### Compatibility

- DSH `0.1.0-rc.6` remains within the declared peer range.
- DSH `0.1.0-rc.7` is the reproducible clean-room and locally integrated
  baseline.

### Safety

- A preview apply acquires the operation lock, so it cannot race a normal sync.
- Applying a preview never touches the watermark (`.last-sync`); a later normal
  sync still sees the sessions that produced the preview.
- Expired previews are never listed and cannot be applied.

## [0.3.0] - 2026-08-20

### Added

- Host-side operation lock (`<root>/.sync/operation.lock`) that serializes sync
  and rollback operations; stale locks from dead processes are recovered by
  pid/mtime checks.
- Active-run tracking (`<root>/.sync/active-run.json`) with phase progression
  (staging/validating/applying/finalizing/complete); an interrupted run left by
  a dead process is recovered into the journal on the next sync.
- `memory.health()` reporting lock, active-run, interrupted-run, journal
  readability, and a `needsManualRecovery` flag; `memory.runs()` now accepts
  `operation` and `status` filters; `memory.status()` reports the newest
  pending preview.
- Hard staging limits enforced before apply: 1 MiB per file, 50 added files,
  5 MiB total change bytes, with stable rejection codes.
- Failed applies are journaled (`status: failed`, `error_code`) so attempted
  runs are auditable.
- `.sync` directory created by the initializer with a `.gitignore` covering
  `operation.lock`, `active-run.json`, and `previews/`.

### Compatibility

- DSH `0.1.0-rc.6` remains within the declared peer range.
- DSH `0.1.0-rc.7` is the reproducible clean-room and locally integrated
  baseline.
- DSH `rc.8` and later releases are not verified by this release.

### Safety

- The operation lock and active-run record are host-side coordination state,
  never committed into the memory payload; `.sync/.gitignore` keeps them out of
  journal commits.
- Dry-run remains read-only: it never touches the lock, the journal, Git, or
  the watermark.

## [0.2.0] - 2026-08-20

### Added

- Transactional idle-session synchronization through an isolated staging
  repository and host-owned recovery/apply commits.
- Metadata-only sync journals, `memory.runs()`, `memory.rollback()`, and
  recent-run status reporting.
- Structured Markdown front matter with incremental legacy-file migration and
  a dry-run preview that does not modify the live memory repository.
- Settings UI support for inspecting the latest sync and rolling back the
  latest journaled apply.

### Compatibility

- DSH `0.1.0-rc.6` remains within the declared peer range.
- DSH `0.1.0-rc.7` is the reproducible clean-room and locally integrated
  baseline.
- DSH `rc.8` and later releases are not verified by this release.

### Safety

- Existing legacy memory files remain readable without a forced migration.
- Rollback creates a new Git commit and never resets or rewrites history.
- Sync journals contain metadata only; transcripts, prompts, and credentials
  are not persisted by the journal.

## [0.1.0] - 2026-08-19

### Added

- `dsh-memory` host plugin with an immediately registered `memory.enabled` setting.
- `dsh-memory-ui` settings row with repository status, a persisted toggle, and a two-step clear confirmation.
- Git-backed, recovery-commit-based memory clearing with symbolic-link and layout protections.
- Portable local installer that creates a missing private Git memory root, optional idle-session synchronizer, transcript redaction, release guard scripts, and GitHub Actions CI.
- Release guards that inspect the Git index snapshot, reject non-text source candidates, and cover classic and fine-grained GitHub token shapes.

### Compatibility

- DSH runtime peer range `^0.1.0-rc.6`, Node.js 22, and Python 3.11.
- The clean-room development test graph uses DSH client packages
  `0.1.0-rc.7`; npm cannot resolve the registry's `rc.6` transitive peer graph
  with plain `npm ci`.

### Not included

- npm registry publication.
- Remote memory hosting or telemetry.
- A guarantee of privacy erasure from local Git history.