# Changelog

## 0.7.0

- Support the DSH 0.1.5-rc.2 persona-prefix, session projection and browser SessionId APIs.
- Add independent long-term memory reconciliation after successful native compaction, using pre-compaction conversation evidence rather than the generated summary.
- Persist maintenance jobs, process long evidence in ordered batches, retry interrupted calls and check memory revisions before committing.
- Preserve Chat/Work isolation: cross-mode maintenance operations become pending bridge writes.
- Remove shortest-card eviction and increase the per-section capacity from 3 to 100 entries.
- Distinguish model-extracted facts from direct user edits; retain evidence references and change records.
- Replace outfit-specific UI wording with general current state, and remove the automatic outfit reminder injection.
- Fix blank-session inheritance for the new workspace API. Existing V5 data is preserved; older formats retain the migration path.
- Document plugin-wide maintenance settings, per-session compaction settings, runtime compatibility and backup/migration behavior.

## 0.6.10

- Added a one-click **设定继承** action in Memory Center. It saves the current memory, creates a blank conversation in the same workspace, copies both Chat and Work faces plus the neutral bridge and compaction policy, then opens the new conversation without copying chat history.

## 0.6.8

- Added dedicated compact tools for reading and writing the AI's current state.
- Added a turn-local reminder for appearance and outfit corrections so the memory duty remains adjacent to the user message.
- Stopped returning the full audit history from model-facing write tools.

## 0.6.6

- Turned `get_session_memory` into a compact authoritative model inspection tool: it now returns only the active memory face and bridge instead of replaying the entire audit history into model context.
- Required a memory read when the user asks what is stored, says a claimed state is not visible, or disputes a remembered fact; an empty field must be acknowledged and corrected rather than covered with conversational prose.

## 0.6.5

- Made explicit, user-confirmed memory changes an in-turn model duty instead of an optional suggestion, including contextual confirmations such as “this outfit” and “keep it this way”.
- Added first-class `set_assistant_state` and `set_assistant_setting` actions so current appearance and role changes no longer require a misleading generic action.
- Made ordinary memory mutations atomic: the active memory is read server-side, while `get_session_memory` remains available for genuinely complex edits rather than being a mandatory extra tool call.

## 0.6.4

- Added an idle-turn threshold enforcer for DSH Web compositions that keep the standing-preset provider private: it measures the real session pressure and invokes the stock `/compact` command once after a completed turn.
- Applying a lower threshold to an already-idle conversation now performs the same check immediately.

## 0.6.3

- Accept the production DSH compaction service when its optional model-policy table is omitted, while preserving every provider-owned setting during each session-local policy overlay.

## 0.6.2

- Restored the stock `/compact` manual action that was accidentally removed from the V5 Memory Center.
- Added live compaction diagnostics for the selected conversation: estimated pressure, model context capacity, actual trigger budget, capacity-safe retained tail, and the latest automatic/manual result.
- Split compaction policy application from the full memory save so compression controls take effect immediately and report transport failures.
- Restored the summary token limit control and clarified that automatic pressure checks run before model steps.

## 0.6.1

- Split one stable user/AI relationship into task-conditioned Chat and Work memory faces without changing tools or permissions.
- Added model routing and a composer mode chip backed by the same session sidecar.
- Added the neutral transition note and pending cross-domain write queue; inactive-mode facts are reviewed only after entering their target mode.
- Migrated V1–V4 business data into Chat while leaving Work empty, and removed the retired automatic-extraction implementation.
- Refined the Memory Center layout and made the composer chip follow model-side changes across live windows.

## 0.5.0

- Promoted explicit AI requirements to the session-owned DSH persona section, so they replace the preset persona instead of being appended after it.
- An empty AI-requirements section now produces an empty persona; removed the forced new-session onboarding and empty-world prompt text.
- Kept people and ordinary memories in a separate contextual prompt section so identity rules are not duplicated or mixed with remembered context.

## 0.4.1

- Preserved unchanged migrated V3 person text beyond 300 characters while keeping the limit for every newly edited value.

## 0.4.0

- Replaced the single-user V3 document with V4 multi-person memory: up to five ordered people, each with a stable id, name, information, one preference, and a relationship/background with the active AI.
- Person one now represents the current speaker without defining that person as the AI's entire world or permanent principal.
- Replaced the roleplay-only preset with up to three ordinary memory groups; AI requirements remain a separate three-card section.
- Added lossless V1/V2/V3-to-V4 migration, including preservation of grandfathered profile text beyond the new 300-character edit limit.
- Updated the model tools, extraction contract, Chinese prompt fields, strict Host/Remote descriptors, Memory Center UI, tests, and documentation.

## 0.3.2

- Synchronized the generated Host/Remote type metadata with the V3 document schema.

## 0.3.1

- Tightened read-before-write so one memory read authorizes exactly one classified mutation.

## 0.3.0

- Replaced AI observations with revisable pending-confirmation user information.
- Defined explicit boundaries for user profile, preferences, assistant requirements, current relationship state, and user-authored roleplay.
- Required the model to read current session memory before every governed write.
- Removed permanent relationship missions and all identity/persona replacement hooks.
- Added V1/V2-to-V3 migration; legacy observations become pending confirmation and legacy missions remain only as historical context.

## 0.2.35 - 2026-08-22

- Qualified the plugin against DeepSeek Harness `0.1.1-rc.2` and the stable
  `0.1.1` line by raising the peer-dependency floor and removing the RC8 suffix.
- Updated manual compaction for the current three-argument Commands Remote and
  report the settled command result instead of a transport-level false success.
- Made automatic retention model-capacity safe: an impossible absolute tail is
  capped below half of the configured trigger budget, with a ratio fallback
  when catalog capacity is temporarily unavailable.
- Verified the installed Web profile with a real `/compact`: 911 history items
  and approximately 314421 tokens were compacted successfully.

## 0.2.34-rc8 - 2026-08-21

- Fixed context compaction after RC8 moved `compaction-basic` and `/compact`
  into standing Agent presets. The plugin now adapts the provider resolved from
  each live `agent.ctx`, so saved per-session thresholds and summary limits
  control the engine that actually owns the conversation.
- Removed the incorrect host-plane compaction reactivation. Automatic and
  manual compaction now share one serialized, session-scoped policy bridge
  without creating a second engine.

## 0.2.33-rc8 - 2026-08-21

- Aligned Memory Center's selector with the Workspace sidebar: only top-level
  sessions owned by a workspace are eligible, then the global archive set is
  subtracted. Subagent and legacy orphan sessions no longer appear as if they
  were restorable user conversations.

## 0.2.32-rc8 - 2026-08-21

- Fixed a cold-start archive race in Memory Center. The session selector now
  waits for both workspace and session baselines before rendering, so the
  initial empty archive placeholder can never expose archived conversations.
  Archived personalization remains retained for an explicit restore.

## 0.2.31-rc8 - 2026-08-21

- Bumped the sidecar format marker so installations that briefly received the
  first sidecar build re-run the safe, read-only legacy import once. No
  conversation log is rewritten by this migration; repaired originals remain
  alongside timestamped backups.

## 0.2.30-rc8 - 2026-08-21

- Extended sidecar migration to the original Memory Center records
  (`memory/set`, `memory/remove`, `memory/relationship`, and summary-edit
  envelopes) as well as the later governed-memory vocabulary. Existing
  preference cards and relationship missions are retained during the move,
  rather than merely making their conversations replayable.

## 0.2.29-rc8 - 2026-08-21

- Moved all editable personalization state and per-session context-compaction
  policy out of DSH's canonical JSONL event log into an atomic, session-keyed
  sidecar under `DSH_HOME/mindspace-session-memory/v1`. Legacy event-sourced
  state is imported once on first access; future edits never append an
  out-of-tree event that a stock RC8 replay could reject.
- Disabled automatic extraction by default. It remains available as an
  opt-in sidecar writer once explicitly enabled, while normal memory changes
  remain model-tool or user-editor initiated.

## 0.2.28-rc8 - 2026-08-21

- Repaired legacy session-memory audit envelopes: historical extraction events
  written before they were explicitly ignorable can be normalized in place by
  the included repair script, with a full sibling backup before any rewrite.
  This prevents a strict RC8 session replay refusal from making an otherwise
  intact Memory Center document appear empty.
- Memory Center now excludes archived sessions from its selector immediately.
  Archiving is reversible: the underlying session and its personalization stay
  retained for an explicit DSH restore, but cannot be read, edited, or injected
  while archived.

## 0.2.27-rc8 - 2026-08-20

- Fixed session-identity drift in mission-bearing conversations. The scoped
  prompt assembly now replaces the Web deployment's “coding agent” persona
  with a capability-only statement while retaining the user-assigned identity
  as the single identity declaration.
- Strengthened the session identity contract across coding, research,
  administration, planning, and roleplay. Those modes are functions, not
  replacement personas.
- Added a compaction-continuity directive: checkpoints retain transient work
  state and must not restate, reinterpret, downgrade, or overwrite the active
  session identity.
- Clarified the documentation: DSH passes the current system context into its
  cache-aligned summarization call; this plugin directs the checkpoint not to
  persist independent identity and personalization layers.

## 0.2.26-rc8 - 2026-08-20

- Added a clean RC8 compatibility path: the bundle no longer patches absent
  in-tree memory rows, so an RC8 profile starts without stale-row warnings.
- Moved the session-specific compaction policy adapter into this package. It
  applies a temporary, serialized policy to the stock RC8 compaction service;
  no `compaction-basic`, bundle, Remote, or TypeScript source patch is needed
  in the Harness checkout.
- This compatibility line targets DeepSeek Harness `0.1.0-rc.8` and is tested
  from an isolated `DSH_HOME`; it must not be installed together with a legacy
  in-tree Mindspace memory implementation.

## 0.2.25 - 2026-08-20

- Re-enabled DSH's compaction engine and `/compact` command after the Web
  bundle overlay, so the per-session context-compaction policy is executed
  rather than merely saved by the Memory Center UI.
- Restored model autonomy for normal memory writes: the main model now decides
  whether stable information merits a tool call. The low-utilization automatic
  extractor skips a turn that already made a successful memory change.
- Reduced cold-start extraction output pressure: its audit list covers only
  durable proposed updates and permits an empty list for ordinary conversation.

## 0.2.24 - 2026-08-20

- Reframed automatic extraction as a cold-start fallback: it now runs only
  while the editable, persisted session-memory document is below 20% of its
  configured capacity. System prompts, RAG, compaction summaries, evidence,
  and event history are not counted.
- Raised the fallback extraction budget to 6000 tokens. Once the cold-start
  threshold is reached, only the primary model's explicit memory-tool decision
  can write memory.

## 0.2.22 - 2026-08-17

- Aligned the checked-in profile defaults with the runtime limits: three cards
  per editable card section and a 300-character user profile.
- Corrected the documented removal command to run through the official Harness
  checkout, matching installation and startup.

## 0.2.21 - 2026-08-16

- Rebased a stale editor save once onto the latest session document. Only the
  sections the user changed are applied; untouched sections retain concurrent
  extractor changes. A repeated conflict preserves the local draft.

## 0.2.20 - 2026-08-16

- Fixed legacy context-compaction policy replay so incomplete historical policy
  events are normalized before crossing the strict Typert result boundary.
- Memory Center now renders editable personalization data even if its separate
  compression-policy endpoint is temporarily unavailable.
- A stale whole-document save no longer retries an old complete document over
  newer memory.

## 0.2.12 - 2026-08-16

- Fixed duplicate Host Typert registration by using the documented manual
  `ctx.typert.register()` ownership for the package's hand-written strict wire
  schema and removing the competing automatic `./typert` loader export.

## 0.2.11 - 2026-08-16

- Fixed live Web upgrades: each strict generated Remote descriptor now mounts
  independently, so existing `get`/`replace` endpoints cannot block newly
  introduced compaction-policy methods.

## 0.2.10 - 2026-08-16

- Fixed the Memory Center save failure caused by putting context-compaction
  fields inside the versioned `sessionMemory/replace` wire request.
- Moved context-compaction reads and writes to dedicated Remote methods, so
  policy changes persist and take effect for the current session without
  rewriting personalization memory.

## 0.2.9 - 2026-08-16

- Added a per-session context-compaction policy: enable state, trigger ratio,
  retained recent context, and summary limit are stored as durable session events.
- Added an editable settings card with a trigger slider and an explicit
  “compact now” action for the active conversation.
- Automatic compaction reads the session policy without incorporating system,
  tool, relationship, roleplay, or personalization prompt layers into the
  summary.
- Tightened the summarizer instruction to preserve durable conversation facts
  while dropping greetings, repeated material, and tool-log noise.

## 0.2.7

- Make clean DSH Web profile installation self-contained by disabling the
  conflicting in-tree experimental session-memory service and its UI row
  before mounting DSH-memory's own Remote and settings surface.

## 0.2.6

- Show the exact session-identity fragment that is injected ahead of the
  default Harness identity after a relationship mission is saved.
- Clarify in the editor that the roleplay preset is a separate enabled-only
  style layer and cannot override the session mission.

## 0.2.5

- Promote an explicit per-session relationship mission into the first identity
  prompt slot, replacing the generic Harness opener without dropping Web or
  tool instructions.
- Keep the roleplay preset as a separate, enabled-only style layer instead of
  conflating it with the session identity.

## 0.2.4

- Fix the Personalization settings panel on current DSH runtimes by resolving
  the plugin-owned Remote after it is mounted, instead of reading it from a
  guarded dependency context during slot rendering.

## 0.2.3

- Mark plugin-owned session events as ignorable persistence records so an
  external plugin cannot make a session unreadable after a cold restart.
- Add a backed-up, atomic repair utility for logs written by 0.2.2 and older.

## 0.2.1 - 2026-08-14

- Removed compatibility overrides for two experimental in-tree entry ids.
- Made the bundle patch cleanly installable on an unmodified official Web profile without `entry not found` diagnostics.

## 0.2.0 - 2026-08-14

- Replaced the public DSH compaction override with a compact confirmed/inferred user profile.
- Consolidated preferences and assistant rules into at most three categorized cards per section.
- Added category-based merge and explicit conflict replacement without model-supplied item ids.
- Added atomic complete-state extraction with a handled/skipped atom ledger.
- Added visible append/merge/replace/skip activity including evidence, before/after values, and reasons.
- Added V1 event migration while keeping legacy history replayable.
- Repaired duplicate fallback categories during replay so legacy documents cannot block later writes.
- Added a dedicated assistant-identity action for names, nicknames, and relationship-specific titles.

## 0.1.0 - 2026-08-14

- Initial community-plugin release.
- Editable per-session memory center and direct DSH compaction preview.
- Preferences, user facts, assistant instructions, relationship mission, and roleplay preset.
- Model read/write tools, conservative extraction, conflict replacement, and audit history.
- One-time optional personalization question for empty new sessions.
- Prebuilt DSH bundle with self-mounted Web Remote and no upstream source patch.