v0.4.8 Release Notes

Platforms: OpenCode (sidecar plugin), server (shared)
Theme: Sidecar reliability + benchmark alignment
Status: Complete

Claude Code impact: None. Changes 1, 2, 4 are OpenCode-only (roampal.ts). Change 3 (fact tags) is in the shared server but is additive — puts tags back on facts.

Overview

Mirror the benchmark's proven sidecar architecture. Four changes:

1. Remove autoSummarize — eliminates Ollama contention that caused 20% scoring failures
2. Remove storeExchange — sidecar fails = nothing stored, no transcript pollution
3. Restore noun_tags on facts — v0.4.7 incorrectly removed them; benchmark tags all content
4. Split sidecar into 2 LLM calls — match benchmark: score_exchange then extract_facts. Tags handled server-side at store time.

Architecture: Benchmark vs v0.4.8

Benchmark flow per exchange (roampal-labs/benchmark/runner.py):

1. sidecar_score()         → {summary, outcome, memory_scores}     60s timeout, temp 0
2. sidecar_extract_facts() → [atomic facts]                        30s timeout, temp 0
3. strategy.store(summary) → extract_tags() on summary at store    30s timeout, temp 0
4. strategy.store(fact)    → extract_tags() on each fact at store  30s timeout, temp 0

Sequential. No concurrency. No background tasks. Tags extracted per-content at store time, not as a separate sidecar call.

v0.4.8 flow per exchange (roampal.ts session.idle):

1. scoreExchange()    → {summary, outcome, memory_scores}     60s timeout, temp 0
2. extractFacts()     → [atomic facts]                         30s timeout, temp 0
3. Store summary via /stop → server store_working() tags it
4. Store facts via /record-outcome → server store_working() tags each

Same sequence. Same timeouts. Server handles tagging via store_working()'s built-in extract_tags — no extra LLM call from the plugin needed.

Note on contention: API models (DeepSeek, OpenAI, etc.) handle concurrent requests natively — contention is a non-issue. For local models (Ollama, LM Studio), sequential calls with no competing background tasks (autoSummarize removed) means clean queueing with no timeouts.

Changes

1. Remove autoSummarize

Files: roampal/plugins/opencode/roampal.ts, roampal/server/main.py

Why: autoSummarize runs in session.idle and hits Ollama via the server. Ollama processes one request at a time. When autoSummarize runs, scoring requests queue behind it → 30s timeout → circuit breaker trips → sidecar dead. Caused 20% scoring failure rate in v0.4.7.

What to remove:

• autoSummarizeOldMemory() function in roampal.ts
• The autoSummarize call in session.idle
• /api/memory/auto-summarize-one endpoint in server/main.py
• _do_auto_summarize_one() and _auto_summarize_one_memory() in server/main.py
• Background task spawn in stop_hook handler (server/main.py)
• Related tests in test_v036_changes.py (TestAutoSummarizeEndpoint)

Why it's safe: Sidecar produces ~200-350 char summaries. No new oversized memories. autoSummarize returned "no_candidates" in testing — dead code.

2. Remove storeExchange

File: roampal/plugins/opencode/roampal.ts

Why: storeExchange() stores raw user+assistant text to /stop BEFORE scoring. If sidecar fails, the raw transcript persists in the DB. The sidecar already stores its own summary via /stop (with conversation_id, fingerprint, outcome, noun_tags, memory_type). storeExchange() is redundant.

What changed:

• storeExchange() function removed (stored raw text to ChromaDB)
• Replaced with lifecycle_only: true call to /stop — registers exchange in JSONL for scoring prompts and calls set_completed(), but does NOT store raw text to ChromaDB
• Without /stop call, set_completed() never fires → scoringRequired stays false → sidecar never runs

New behavior: Sidecar succeeds → summary stored via sidecar's own /stop call. Sidecar fails → nothing in ChromaDB (JSONL still has the exchange for scoring prompts).

3. Restore noun_tags on facts

File: roampal/server/main.py (shared — affects both clients)

Why: Benchmark entity_routed.py:248-250 extracts LLM tags for ALL content at store time. v0.4.7 removed tags from facts. Tagged facts participate in TagCascade retrieval.

Changes:

1. Remove the noun_tags=None override on fact storage — let store_working() handle tagging via its built-in extract_tags (same as benchmark: tags extracted from fact content at store time)
2. Remove v0.4.7 tag-stripping migration (server/main.py:524-548)
3. Add v0.4.8 re-tag migration: LLM extraction via sidecar extract_tags() for facts with noun_tags="[]", regex fallback if sidecar unavailable

4. Split sidecar into 2 LLM calls

File: roampal/plugins/opencode/roampal.ts

Why: Matches benchmark architecture. Each call has a focused prompt → better quality from small models. If one fails, the other can still succeed.

Current (v0.4.7): One prompt asks for {summary, outcome, noun_tags, facts, memory_scores}.

New (v0.4.8): Two sequential calls:

1. scoreExchange() → {summary, outcome, memory_scores} — 60s timeout, temp 0
2. extractFacts() → [fact1, fact2, ...] — 30s timeout, temp 0

Tags are NOT extracted by the plugin. The server's store_working() extracts tags from each piece of content at store time — matching exactly how the benchmark's entity_routed.py:store() calls extract_tags() internally.

Noun tags for the exchange summary come from the server-side tagging when the summary is stored via /stop. No separate plugin LLM call needed.

Benchmark Reference

| Function | File | Timeout | Returns |
|---|---|---|---|
| sidecar_score() | runner.py:221 | 60s | summary, outcome, memory_scores |
| sidecar_extract_facts() | runner.py:173 | 30s | list of atomic facts |
| extract_tags() | entity_routed.py:74 | 30s | list of noun tags (called at store time) |

All use temperature: 0. No retries. Exceptions caught and skipped.

Expected Impact

• Scoring reliability: ~100% (no contention, sequential calls)
• Scoring quality: Better (focused prompts vs overloaded combined prompt)
• Fact retrieval: Improved (tags enable TagCascade routing)
• DB cleanliness: No raw transcripts
• Complexity: Reduced (~200 lines autoSummarize removed)

Verification

1. No autoSummarize in plugin debug log
2. No Stored exchange in plugin debug log (only SUMMARY stored)
3. 2 separate LLM calls per exchange visible in debug log (score → facts)
4. Facts stored with noun_tags (check working collection metadata)
5. 10+ exchanges with 0 circuit breaker trips
6. Re-tag migration: facts with noun_tags="[]" now have tags
7. Claude Code: score_memories still works, facts get tags

Tests

• Remove TestAutoSummarizeEndpoint tests
• Add: sidecar failure → no exchange stored
• Add: facts stored with noun_tags
• Add: 2-call sidecar flow (score → facts)
• Add: re-tag migration
• Verify Claude Code MCP tools unaffected