v0.17.0 — Flywheel & Project

"Flywheel & Project" (M19)

A capability release (not a benchmark-headline one — see the honest note at the end).

1. The self-evolution flywheel now turns on every autonomous path — and reads back what it learns

The machinery already existed (learned skills → cards, GEPA prompt evolution, an ACE playbook, long-term memory, a diff-gate, a measured skill lifecycle) but was wired only inside solve, and mostly write-only.

• A shared EvolutionContext makes learning a property of the agent stack, so the kanban lanes, workflow steps, and the SDLC lifecycle crew all learn — not just solve.
• Memory is read back into runs (it used to only be written).
• The diff-gate stops "hollow successes" (verifier passed, but nothing actually changed) from minting skills or memories — you only learn from work that changed something.
• The hierarchical orchestrator reads recalled facts into its synthesis and records outcomes, without distilling skills from fan-outs (which have no verify-or-revert signal).

2. chimera project — run a whole project start-to-finish against a Spec

A drift Spec (a small YAML of requirements) is the executable "done" authority — nothing is accepted on the model's say-so. Each unsatisfied requirement becomes a card verified by chimera drift --only <id>; a Kanban board tracks dependencies (with a blocked column); the loop runs until the spec aligns or a rail stops it: max-iterations, a high-risk step awaiting your approval (deploy/migration/delete pauses), or a stuck card that escalates to a human. Durable and resumable. project start / status / run / step / approve / deny.

3. Closing the loop: evolve refine + evolve guard

• evolve refine mines verified runs to GEPA-refine a skill, but only adopts it if it passes a transfer gate (helps its tuned slice and doesn't regress a held-out same-capability slice — no holdout ⇒ dry-run).
• evolve guard auto-rolls-back the most recent skill on a statistically significant regression (a confidence interval, never a point estimate), via a reversible retire.

Honest scope note

The one change that would need a benchmark to justify — reading learned skills by default — ships wired but OFF, gated on a paired skillcard-bench showing Δ ≥ 0. No flipping defaults by faith.

Install: pip install -U chimera-agent · Apache-2.0