v0.9.3 — Smaller README, Evidence column, runnable Quick Start

• README landing page is smaller (27 KB → 21 KB): the 15 case-study <details> blocks are now one table row each, with the full write-up in examples/.
• README "Real-World Results" and "Tooling Measured Results" gain an Evidence column (measured, estimated, single run, qualitative). The unsourced "100x faster", "100% tool consistency" and "100+ skills" claims are reworded or marked as observed, not measured, in the README, skill-tooling-pattern.md and examples/index.md.
• Quick Start is now three runnable steps built on claude-md-lint.py, replacing a grep of private paths and two comment-only steps.
• BACKLOG.md: re-review adds seven items (B-007, B-008, B-017, B-018, B-027, B-028, B-029) and reprioritizes; five items shipped. GitHub topics and homepage are now set.

v0.9.2 — macOS CI and anydoc error

• CI now runs every check on macos-latest as well as Ubuntu, so the macOS branches in loop-detect and statusline are tested.
• tools/doc-convert/anydoc_to_md.py now exits with the uv run --with firecrawl-anydoc command when anydoc is missing, instead of a ModuleNotFoundError traceback. Added a --self-test for that message, run in CI. It does not run a real AnyDoc conversion; the README says so.

v0.9.1 — BACKLOG.md

• Added BACKLOG.md — a living list of planned updates and improvements (33 items from a full review at v0.9.0, with IDs, priorities and maintenance rules), linked from the README and CONTRIBUTING.

v0.9.0 — loop-detect retry-loop guard

• Added tools/loop-detect — a PreToolUse hook that asks for confirmation on the 3rd identical tool call in a row, to catch retry loops (bash plus jq). No measured savings: zero such loops in the author's 13,753-call history. Includes a --self-test, run in CI.

v0.8.1 — Figures redrawn as SVG and Mermaid

• Replaced the nine AI-generated PNGs in resources/ (11.5 MB) with seven hand-built SVGs (48 KB) on one dark-card style, plus two Mermaid diagrams for the decision tree and before/after flow. Text is now real text, so a number change is an edit, not a regeneration. Fixed the typo ("Retrieval"), garbled labels and other-vendor logos in the old images, and removed the duplicated "Visual Reference" table. ContextSurface and the cover image now have alt text that states their point.

v0.8.0 — retrieval-gaps

• Added tools/retrieval-gaps — mines your own Claude Code transcripts for the times you had to point Claude at the right file ("it's in the X wiki", "you missed that"), so you can count real retrieval misses before building retrieval infrastructure. Python stdlib, local, no model tokens; includes a --self-test, run in CI. Extracted from the new Evaluate Before Adopting case study.

v0.7.0 — skill-usage

• Added tools/skill-usage — silent PostToolUse hook that logs one JSONL line per skill call, plus a stdlib Python report of calls, last-used date, and installed skills never or rarely used against the size of their always-loaded listing entry (measured in characters, not tokens). Extracted from the Skill Tooling Pattern setup. Includes a --self-test.

v0.6.0 — cost-by-task, derived-stale and statusline tools

• Added tools/statusline — the context-size status line from the Context-Length Threshold case study as a standalone script (bash plus jq): grey, amber at 60% of the compact window, red at or over it. Includes a --self-test, run in CI.
• Added tools/cost-by-task — Claude Code spend per task id, joined from your own transcripts and reconciled to ccusage (Python stdlib plus the ccusage CLI). Extracted from the Cost Attribution to Tasks case study; needs a task tracker exposed as MCP tools (set --tool-prefix and --id-field). Includes a --self-test, run in CI.
• Added tools/derived-stale — content-hash staleness check for derived documents (Python stdlib, no dependencies): sidecar hash, 5% size materiality, a thin-source guard that keeps a .prev backup, and a peer-median thin check. Extracted from the Stale Detection case study (output lint and weekly nudge not included). Includes a --self-test, run in CI.

v0.5.2 — Status line and ccusage audit recipe

• Extended the Context-Length Threshold case study with a portable status line script (context size against the auto-compact window, colored as it nears it) and a ccusage before/after audit recipe, including the pitfalls hit while measuring: ccusage reprices history, a cost-per-day drop is confounded by workload and model mix, and a cap on context does nothing for sessions that are long in turns rather than large in context.

v0.5.1 — Incremental Wiki Refresh case study

What's Changed

• docs: routing-regression findings for context trim case study by @zwolf25 in #38
• docs: auto-compact window follow-up to context-length case study by @zwolf25 in #39
• v0.5.1: Incremental Wiki Refresh case study by @zwolf25 in #40

Full Changelog: v0.5.0...v0.5.1