v0.8.0

Fixed

• ai-forge-judge: the description colon rule was wrong. It auto-failed S1 on any colon, but a colon inside a quoted scalar (description: "… animate:flip …") is valid YAML and the agentskills spec imposes no colon rule of its own. Only a bare key: value shape in an unquoted value breaks parsing, and only that is now an auto-fail. The old rule mis-graded correct skills.
• ai-forge-judge, ai-forge-review: evals/ and benchmarks/ were flagged as skill-root spec violations. A purpose-named directory is not a stray file; only non-SKILL.md files at the root are. Both now say so explicitly, so an eval suite no longer costs a skill points.
• ai-forge-update: Phase 3 invoked ai-forge-apply unconditionally while the judge's report already ended in its own (y)es / (n)o prompt, so the user was asked twice. Phase 3 now honours that prompt and defines the (n) branch.
• ai-forge-eval: an agent's result and its <usage> block can arrive in two separate messages. A hand-back without usage was being treated as permanently missing data; it is now recorded as null and overwritten if a later notification for the same task id carries the block.

Changed

• ai-forge-apply: one decision per turn. The uncommitted-changes warning, the git opt-in, the branch question, and (A)/(s) are four turns rather than one line — a human facing two prompts answers both in one reply ("n A"), which the skill had no defined parse for. Added as a NEVER with that rationale.
• ai-forge-apply: the git branch decision moved to references/git.md, loaded only on (c)ommit. It also gains an escape hatch — a repo whose own convention puts AI work on the default branch commits there instead of forcing an ai/* branch.
• ai-forge-apply: (r)evise now collects one sentence about what was wrong with the change, not with the finding.
• ai-forge-judge: dimension U3 renamed Anti-Pattern Quality → Constraint Quality; the report footer is now the literal prompt Step through these with ai-forge-apply? (y)es / (n)o.
• ai-forge-eval: trigger validation runs each query twice, with a third run only where the two disagree. A measured 12-query set returned unanimous judgments on all 24 runs, so a blanket third run mostly re-buys an answer already in hand.
• ai-forge-create: benchmarks/ added to the directory taxonomy for committed trend baselines.

Added

• ai-forge-apply: references/git.md.
• ai-forge-eval: agents/refiner.md now states that it has not read the artifact, so every improvement must name a behaviour visible in the outputs — never structure, ordering, or prominence. A caller who knows the file discards the whole list when one item is invented.

v0.7.0

Fixed

• ai-forge-apply: description advertised "commits after each approved item", contradicting the body's NEVER rules, which forbid per-item commits and require a single opt-in commit after the loop. The description is the only text the router sees, so the contradiction was user-visible.
• ai-forge-update: NEVER #5 ("never proceed to Phase 3 if apply applied zero changes") read as forbidding the zero-findings branch that Phase 3 itself defines. Scoped the rule to Phase 2's apply and stated that it does not govern Phase 3's own exit.

Changed

• ai-forge-create, ai-forge-eval, ai-forge-recap, ai-forge-review, ai-forge-update: descriptions rewritten as trigger conditions rather than workflow summaries, so the router cannot act on the description in place of loading the body.
• ai-forge-create, ai-forge-eval: added failure branches for their validator scripts. A validator that runs and rejects still blocks; a validator that cannot run (missing node, missing script, usage error) now degrades with a recorded ⚠ unvalidated caveat instead of being an undefined state. Both cases currently exit 1, so the branches key on the error text — giving the scripts a distinct exit code remains open.
• ai-forge-audit: already had a failure path for check-ecosystem.cjs; it now stamps ⚠ roster-level checks skipped on the report rather than only noting the skip, since the six roster checks are precisely what per-artifact grades cannot see.
• ai-forge-create: assets/SKILL.template.md is now referenced from the body; it was previously unreachable.

Full Changelog: v0.6.2...v0.7.0

v0.6.2

Security (v0.6.1)

• ai-forge-judge: documented that WebFetch results (e.g. the agentskills.io spec) must be treated as inert reference text, never as instructions — mitigates indirect prompt injection risk (skills.sh audit finding W011).
• ai-forge-audit: documented that check-ecosystem.cjs is local and read-only — no network access, no subprocess spawning, no writes outside stdout — addressing a "delegated execution / transitive trust" flag from the skills.sh audit.

Changed (v0.6.2)

• ai-forge-audit, ai-forge-create, ai-forge-eval: reformatted check-ecosystem.cjs, validate-metadata.cjs, and aggregate-benchmark.cjs to single-quote/no-semicolon-trailing-comma style with lint-disable annotations; no behavioral changes.

Full Changelog: v0.6.0...v0.6.2

v0.6.0

Added

• ai-forge-create: Phase 0 dedup check and Phase 1b baseline probe — measure what the agent does without the artifact before drafting, so observed failures drive the body instead of asserted ones. New references/baseline-probe.md carries the protocol and a failure-form matching table (guidance shape follows failure type).
• ai-forge-create: references/drafting-craft.md for leading words, degrees of freedom, completion criteria, and the constraint format.
• ai-forge-eval: Phase 0 loads or writes a persisted evals/evals.json, so a suite survives the conversation and can be re-run as a regression test. New references/eval-suite.md documents the format, trigger protocol, and pressure escalation.
• ai-forge-eval: Phase 3b human review step with a self-contained assets/eval-review.html template; Phase 6 trigger check reporting recall and precision against a 60/40 train/held-out split.
• ai-forge-eval: assertion discrimination — classifies each assertion by whether it distinguishes the with-artifact arm from baseline, so a suite of assertions that pass either way stops reading as a healthy pass rate.
• ai-forge-audit: Phase 0 ecosystem coherence pass and scripts/check-ecosystem.cjs — trigger collisions, duplicate and shadowed names, stale links, orphaned references, dated model pins, and word budgets across a whole roster.
• ai-forge-review: refutation pass with a severity-ordered charge and blocker/should-fix/note labels.
• ai-forge-update: references/iteration-guide.md with iterate-vs-redesign sizing and a signal-to-fix table; conversion now runs a field-disposition report and routes judgment calls through ai-forge-apply.
• ai-forge-create: validate-metadata.cjs gains --target (per-platform disposition report), --when-to-use, a dated-model-ID check, and the 1,536-char combined listing cap.
• Documentation for the full Claude Code SKILL.md frontmatter field set, skill-scoped hooks (including the exit-code-1-does-not-block trap), the script-vs-prose decision, and domain-variant reference organization.

Changed

• ai-forge-judge: U3 renamed to Constraint Quality and rescored. It now grades whether constraints are justified, explained, and load-bearing rather than whether a NEVER list exists — a constraint carried by explained reasoning scores higher than the same constraint asserted, and prohibition bloat is penalized.
• Reference-file TOC threshold raised from 100 to 300 lines, matching the spec.
• Skill body budget stated as 200 lines and ~1,800 words, set to bind at the same point.

Fixed

• Three shipped skills failed this pack's own metadata validator (unescaped colon in description).
• ai-forge-review/assets/AI-SPEC-TEMPLATE.md carried source-project specifics; generalized, and given requirement traceability and a decisions table.
• AGENTS.md leakage check widened from a single project-name grep to cover ticket IDs, absolute home paths, and brand names, plus self-conformance checks.

v0.5.0

Initial release: ai-forge-apply, ai-forge-audit, ai-forge-create, ai-forge-eval, ai-forge-judge, ai-forge-recap, ai-forge-review, and ai-forge-update skills.