v0.3.0

Added

• Step 4 — Contest the findings, with references/contest.md. Every finding from Steps 2 and 3 is a hypothesis until it passes evidence (cite the moment), counter-evidence (find the moment that contradicts it; a rule founded on fewer than three moments is a hunch), and mechanism (would the fix have changed the cited moment, and does it conflict with a sibling skill's rule). Verdict per finding: confirmed, reframed, or falls — only survivors reach the Wiring section. A first pass here was two-of-seven wrong and missed the strongest item, which is what the step exists to catch.
• Step 1 — "Was it seen?": before the score is final, list every user-facing surface the session changed and who looked at it — a capture, a run, a person, or nobody. A surface nobody looked at is a finding whatever the tests said.
• A score cap of 6 for any session that shipped an unseen surface, overriding the scoring table in references/shapes.md. One session scored 7/10 on context economy alone and came back with nine defects from a single play.
• Contested section in the output — finding, verdict, and the moment it stood or fell on.

Changed

• "Write the notes" renumbered Step 4 → Step 5; Do NOT Load updated to match.
• Step 2: magnitude stays relative except where the session's own tool output printed a number (a suite runtime, a git log count) — cite it, with its source.
• Step 3: an unseen surface gets its lever from the gate of the artifact that shipped it, never a note.

v0.2.0

Added

• Environment levers reference (references/environment-levers.md): six named levers — navigation pointer, automated check, review-stage rule, steering trim, tool verbosity, information access — each with a use-when signal, its natural home file, and a severity ordering. Step 3 now asks "turn or ground?" for each badly-spent moment before asking what was done by hand around a called artifact, so a cost that would recur for any agent gets a fix at its source instead of a "be more careful" note.

Changed

• Wire-in note gains a lever field and may target a steering file, a check config, a review-standards file, or a tool wrapper — not only a skill or workflow.
• Context-economy reference now routes ground-side costs (verbose-by-design tools, unreachable facts, hard-to-find files) to the levers rather than stretching a turn-side rule over them.
• Wiring output lists recommendations most severe first.
• Description names the lever concept so the router can tell this skill proposes environment fixes, not only skill edits.
• Eval suite moved from skills/retrospective/evals/ to evals/retrospective/ so the skill directory holds only SKILL.md and references/, per the Agent Skills spec.

v0.1.0

Full Changelog: https://github.com/robcsaszar/retrospective/commits/v0.1.0