v0.2.0

Correctness and context. Twelve defects are fixed — including one that made a correct harness fail the skill's own delivery gate — and the context a session pays for before its first prompt is cut in half. Every product mechanic asserted by the skill was re-verified against live documentation, because the previous release shipped one that was wrong.

Breaking

• audit_harness.py --json: inventory.claude_md is now an array, not an object. A project CLAUDE.md can live at ./CLAUDE.md or ./.claude/CLAUDE.md, and ./CLAUDE.local.md loads alongside either, so a single slot could not represent reality. If you read inventory.claude_md.lines, read inventory.claude_md[0].lines — or iterate, which is the point of the change. An absent CLAUDE.md is now [] rather than null.

Fixed

• @import detection no longer rejects correct harnesses. maintainer: ops@acme.com was read as an import of acme.com and react@18.2.0 as an import of 18.2.0, each raising an error and a non-zero exit — so a CLAUDE.md that named a maintainer or pinned a version could never satisfy validate_harness.py. Import parsing now skips fenced blocks and code spans, requires a path boundary, and supports the documented extensionless @README form. Relative targets resolve against the importing file rather than the repo root, and @~/… imports are recognized as external rather than reported missing.
• A project using .claude/CLAUDE.md is no longer reported as having no harness. Both scripts hardcoded ./CLAUDE.md, which fed the mode suggestion and classified an established harness as new.
• Rules and agents are discovered recursively. A nested .claude/rules/frontend/style.md without paths: loads at launch but was invisible to the linter, the inventory, and the drift check.
• Spec drift is detected in both directions. The audit only reported components on disk that the spec omitted; it now also reports spec rows whose status claims a file that isn't there, and in_spec_not_on_disk is present in --json whether or not a spec exists.
• The workflow syntax check accepts top-level return, which the workflow runtime supports and which the skill's own examples use — it was reported as a syntax error.
• SessionEnd's 1.5-second timeout is documented correctly as a budget shared across all SessionEnd hooks, not a per-hook default; two reference files also disagreed about which event had the shortest timeout.
• Wrap-up now validates after the edits it is meant to check, rather than before them.
• Dead-link checking covers a skill's pointers written as prose, markdown links, or ${CLAUDE_SKILL_DIR} invocations — previously only backtick-wrapped forms in one file, one pointer out of thirty-two.
• Three dangling cross-references and an unclosed code fence.

Added

• An always-loaded budget report, printed on every validate_harness.py run: CLAUDE.md plus expanded @imports plus every rule without paths:, with an explicit list of the surfaces it cannot count (user scope, ancestor directories, auto memory, managed policy).
• Three lint warnings, each with a fixture proving it stays quiet on correct input: generic advice anchored to whole sentences, a deny rule that swallows an allow rule, and a catch-all paths: glob.
• references/re-entry.md — the extend/improve/sync modes and the full drift-resolution procedure, loaded only when re-entering an existing harness.
• Auto memory and the personal-vs-team scope axis. The routing framework now asks who needs a fact and who writes it, with CLAUDE.local.md as the destination for per-developer facts; auto memory enters the budget model but is explicitly never a routing destination, being nondeterministic and disableable.
• Mechanics whose absence leaves a generated harness silently inert: what workspace trust gates beyond allow rules, why Edit(.claude/**) cannot pre-approve a protected-path write, that defaultMode: "auto" is ignored in project settings, that Write(path) permission rules are never consulted, and that enabling a subagent's memory re-enables Read/Write/Edit.
• A compaction survival matrix — root CLAUDE.md and unscoped rules are re-injected, subdirectory CLAUDE.md and paths:-scoped rules are not — which qualifies the monorepo guidance in the same file.
• declined and retired spec statuses, so a harness records what was deliberately not built.
• This repository now carries its own .claude/harness-spec.md.

Changed

• The always-loaded surface is 4,833 → 2,411 words. SKILL.md instructed an unconditional load of interview.md, so the split between them bought nothing; Phase 0 now branches by mode and the sync path never opens the interview file.
• Principles are stated once rather than elaborated, and examples no longer carry domain narrative that pulled generated components toward the example's subject matter.

Known limitations

• Headless authentication propagated correctly during this release's verification, but run_e2e.py's permission handling remains a documented best guess rather than a broad confirmation.
• The interview still cannot be end-to-end tested; AskUserQuestion does not exist in headless or subagent contexts.
• A generated harness does not yet mention that workflow agents run in acceptEdits mode and inherit the session allowlist, though the reference prose covers it.