Skip to content

v0.2.1 — Dialectic Stage 2 parse tolerance

Latest

Choose a tag to compare

@JuzzyDee JuzzyDee released this 19 May 19:32
df40160

A patch release that hardens the dialectic against incomplete LLM output. Existing v0.2.0 installs will see an update-available prompt on next recall.

What's fixed

The dialectic's Stage 2 dialogue and synthesis steps were silently dropping candidates when Haiku occasionally omitted required fields (evidence on a challenger turn, action_payload on synthesis). Tolerant deserialization + the
existing validation gate now record the attempt properly, apply cooldown normally, and surface incomplete LLM output as validation_failed audit rows you can investigate, instead of having them disappear into errors_summary as parse
errors.

Impact

  • No memory corruption from the prior bug — failed candidates were dropped, not mis-reframed.
  • No user-facing MCP tool was affected (recall, remember, reframe, etc. all worked correctly throughout).
  • This release restores Stage 2 to recording every dialectic attempt, even ones the LLM produced badly.

How to update

```bash
git pull
wrangler deploy
```

No migration, no schema changes, no config changes.

Not in this release

A separate issue where the synthesizer reaches a reframe verdict without producing the new content — also caught by the validation gate, but a prompt-design fix rather than a code fix. Tracked for the next prompt-tuning pass.