OpenCode v1.2.25+ validates that all part IDs must start with 'prt'.
The chat.message hook was creating parts with IDs like
'supermemory-nudge-xxx' and 'supermemory-context-xxx', causing
ZodError crashes on every session start.

Changes:
- Extract generatePartId/generateMessageId into shared ids.ts
- Use generatePartId() in index.ts for nudge and context parts
- Remove duplicated functions from compaction.ts

Fixes #28
Fixes #29
Fixes #32

OpenCode v1.2.25+ validates that all part IDs must start with 'prt'.
The chat.message hook was creating parts with IDs like
'supermemory-nudge-xxx' and 'supermemory-context-xxx', causing
ZodError crashes on every session start.

Changes:
- Extract generatePartId/generateMessageId into shared ids.ts
- Use generatePartId() in index.ts for nudge and context parts
- Remove duplicated functions from compaction.ts

Fixes #28
Fixes #29
Fixes #32

1. Japanese memory keywords:
   Added patterns like 覚えて, メモして, 保存して, 忘れないで etc.
   so the plugin recognizes Japanese memory trigger phrases.

2. Session-end auto-save:
   - Track user and assistant messages during the session
   - On session.deleted event, ingest the conversation into
     both project and user scopes via supermemory
   - Minimum 3 messages required to trigger save
   - Capped at 50 tracked messages / 50k chars to avoid
     unbounded memory growth

fix: use prt_ prefix for all part IDs (OpenCode v1.2.25+ compat)

…on-save

feat: add Japanese keyword patterns and session-end auto-save

Previously, supermemory context was only injected on the first message
of a session. This meant topic changes mid-session lost relevant
memory context.

Changes:
- Add reinjectEveryN config (default 0 = first only)
  Set to e.g. 10 to re-inject context every 10 messages
- Add recall keyword patterns (English + Japanese)
  'recall', '思い出して', 'メモリ確認' etc. trigger immediate
  context re-injection on that message
- Clean up session state on session.deleted
- Track per-session message counts instead of boolean set

Config example (supermemory.jsonc):
  { "reinjectEveryN": 10 }

feat: context re-injection with periodic and keyword triggers

…e, repo tag, signal extraction

Ports all major features from claude-supermemory to opencode-supermemory-max:

1. Entity Context: PERSONAL_ENTITY_CONTEXT and REPO_ENTITY_CONTEXT
   constants guide supermemory on what to extract from conversations
   (user intent, decisions, preferences vs architecture, patterns)

2. Dedup: deduplicate search results and profile facts to avoid
   showing the same memory multiple times

3. Relative Time: show '3hrs ago', '2d ago' etc. alongside similarity
   percentages for memory results with timestamps

4. Repo Container Tag: 3-tier memory scopes (user/project/repo)
   - repo tag based on git remote name for team-shared knowledge
   - search, add, list all support scope='repo'
   - session-end saves to all 3 scopes

5. Signal Extraction: importance-based filtering for session saves
   - only saves turns containing signal keywords (configurable)
   - includes N context turns before each signal for continuity
   - English + Japanese signal keywords built-in
   - disabled by default (set signalExtraction: true to enable)

feat: claude-supermemory parity (entity context, dedup, relative time, repo tag, signal extraction)

- Full English README with feature comparison table vs upstream
- Japanese README (README.ja.md) with complete documentation
- Clear installation, configuration, and feature documentation
- Highlights all max-specific features: Japanese keywords, signal
  extraction, repo tags, dedup, entity context, relative time

docs: comprehensive README (EN + JA)

Ported from OpenClaw supermemory-auto best practices:

1. Incremental capture (crash-safe):
   - Save each user+assistant exchange immediately after assistant responds
   - No more waiting for session end — if OpenCode crashes, memories are safe
   - Uses PERSONAL_ENTITY_CONTEXT for extraction guidance
   - Configurable: incrementalCapture (default: true), maxCaptureChars (5000)

2. Every-message recall:
   - Changed reinjectEveryN default from 0 to 1
   - Context is now re-injected on EVERY message, not just the first
   - Ensures topic changes mid-session get fresh relevant context

3. Metadata stripping:
   - Strip supermemory-context blocks from queries before recall
   - Strip JSON metadata blocks and untrusted sender info
   - Cleaner queries = better recall accuracy
   - Query trimmed to 300 chars max

feat: incremental capture, every-message recall, metadata stripping

1. Pre-compaction full conversation save:
   - Before compaction, capture ALL conversation messages to supermemory
   - Saves to project, user, and repo scopes with PERSONAL_ENTITY_CONTEXT
   - Ensures no conversation data is lost during compaction
   - Allows 2x maxCaptureChars for compaction saves

2. Post-compaction memory re-injection:
   - After compaction, re-inject supermemory context in the 'Continue' message
   - Recalls user profile, user memories, and repo memories
   - Agent maintains continuity across compaction boundaries
   - Prompt instructs agent to use memory context to resume

3. Summary saves to all scopes:
   - Compaction summary now saved to project + user + repo scopes
   - Uses PERSONAL_ENTITY_CONTEXT for extraction guidance

4. Repo tag support in compaction:
   - createCompactionHook now accepts { user, project, repo } tags

feat: pre-compaction save + post-compaction memory re-injection

- Merged upstream/main (supermemoryai/opencode-supermemory v2.0.5)
- Resolved conflicts in package.json and src/index.ts
- Adopted upstream prt_ prefix ID format (removed custom generatePartId)
- Retained all max features: Japanese keywords, signal extraction,
  incremental capture, entity context, session auto-save, compaction hooks
- Version bumped to 2.1.0