v0.5.11

What's new

Progressive compression

Compressed conversation views for conductor and lenses without modifying the raw transcript. The conductor gets a territory-focused digest optimized for gap detection; lenses get a phrasing-focused digest preserving breakthrough quotes and live metaphors. Reduces data sent by ~60% across a 35-turn session. Disable with --no-progressive.

Two-phase conductor

The conductor now operates in two phases per turn: pre-lens (pick which lens speaks next) and post-lens (react to what was actually said). This lets the conductor respond to surprises rather than predict outcomes.

Multi-step lens calls (--multistep)

Experimental. Tool-enabled lenses (Observer, Empath, Skeptic, Mortal, Logician) get a two-step deliberation: explore with tools, then respond without them. Gives lenses autonomy-of-thought without autonomy-of-action.

Dialogue output type (--dialogue)

New output type producing first-person internal monologue. Use --dialogue to add it alongside defaults, or --type dialogue for standalone.

Provocation verification

Provocations with quantitative claims are now fact-checked via web search before sessions start. Fabricated stats get corrected, rhetorical framing gets flagged.

Line prosecution

The winning platform + expression pair passes through an adversarial prosecution loop testing briefability, cold-readability, and specificity.

Other

• --no-context flag to skip project context gathering
• Mechanism memory prevents amnesiac re-discovery across mechanisms
• Conductor can pause for user questions (next_lens: "user")
• Conductor reasoning field logged to log.jsonl and visible in session diagram
• Session diagram visualization (report/session-diagram.py)
• "Cap" terminology renamed to "rate limit" throughout
• Harness and diagram improvements