roampal-ai/roampal-core

Hotfix for v0.5.3 — facts and summary extraction broken on thinking models

Three bugs in the OpenCode plugin's JSON extraction logic. All three only manifested with reasoning/thinking models on local sidecar setups.

What was broken

Bug 1 — Facts extraction silently failing on every exchange (latent since v0.4.8). The facts extraction path used a greedy regex that swallowed trailing markdown code fences. Models that wrap JSON in ```json ... ``` blocks broke JSON.parse() with Unrecognized token '\``'. Errors were logged as non-fatal so it was silently invisible — no facts ever landed in memory for fence-wrapping models.

Bug 2 — Summary/scoring path nested-brace fragility (latent). The summary path used a non-greedy single-match regex which stops at the first }. On responses with nested memory_scores: {...} objects or strings containing }, it captured a fragment and JSON.parse() failed with Expected '}'. Latent until per-memory scoring started returning nested objects in real exchanges.

Bug 3 — max_tokens too low for verbose reasoning models. Scoring max_tokens=2000 was sometimes insufficient for verbose summaries — output got truncated mid-string with no closing brace anywhere for any regex to find.

Fix

• New extractJson() helper in the plugin — tries whole-text parse → greedy match → non-greedy multi-match validation. Handles backtick fences, nested objects, embedded brace characters in one place.
• Both lanes (scoring + facts) refactored onto the helper.
• max_tokens bumped: scoring 2000 → 4000, facts 1000 → 2000.

Verified end-to-end

Manual testing with a thinking sidecar model on LM Studio confirmed multiple consecutive exchanges with SUCCESS + SUMMARY stored + FACTS extracted + per-memory scoring, zero JSON Parse errors.

Sidecar model recommendation

Prefer a non-thinking sidecar model. Reasoning-by-default models emit their chain-of-thought into a separate reasoning_content field while the structured answer goes in content. With longer summaries the model sometimes runs over its token budget mid-reason, leaving content truncated or empty. The hotfix gives thinking models more room to finish and the parser recovers from most partial outputs — but model-side truncation ultimately can't be fixed in the plugin. Pick a non-thinking model in roampal sidecar setup to eliminate this class of failure entirely.

Plus: CI test isolation cleanup

Bundled with this release: a separate test-only commit consolidating Linux + Python 3.10 isolation fixes — XDG_CONFIG_HOME env-var leakage in cli tests, and a skipif for one async dedup test that hangs only on Linux + Python 3.10 (unrelated asyncio quirk; production code unchanged).

Migration

None required. Just pip install --upgrade roampal==0.5.3.1 and restart any running OpenCode session so the plugin reloads.