Bump 0.8.13 -> 0.8.45. Extract the SKILL.md monolith (~530 lines) into
references/ for progressive disclosure: github-and-merge, transcribe,
extraction-spec, exports, update, query, add-watch, hooks. SKILL.md now
points to each reference and loads it only on the path that needs it.
Inline fixes carried by the new version: empty-extraction guard before
any write (#1392), shrink-guard ordering so GRAPH_REPORT/analysis never
describe a graph.json that was refused (#479), root= relativization for
build/manifest parity across clones (#1361/#1417), stale-cache cleanup
and code-only semantic pre-write (#1392), edge-direction preserving
merge (#801). Adds FalkorDB export (--falkordb/--falkordb-push) and
rewrites the frontmatter description (drops the obsolete trigger: field).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0169vjUD1sP9Nx4ZiCa8wvAw
The gate's remedy is now /profile design (not an atomic ui-ux-pro-max toggle),
but several callers still described the old "ask user to activate ui-ux-pro-max"
behavior. Repoint the prose — the `Follow lib/design-gate.md` pointers stayed
valid (the gate was extended in place), so only descriptions change:
- agents/feater.md, hotfixer.md, bugfixer.md (STEP 0.5/1.5 gate bullet):
signals found → run design-tool-gate.sh; if INCOMPLETE → /profile design.
- agents/feater.md rule 6: trigger is "design toolchain incomplete", not
"ui-ux-pro-max inactive".
- agents/plugin-advisor.md: gate points at /profile design when the toolchain
is incomplete.
- skills/feat/test-prompts.json (id 2): expected behavior updated; still valid JSON.
No logic change — descriptions only, aligned with the profile-based gate.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
list: ITEMS column with compact per-category counts (e.g. 12s·1p·1m·1c),
left block aligned; full descriptions kept (CLI output, not 80-strict).
show: grouped by type (gstack/external/personal/plugin/mcp/cli) with
status; redundant per-line TYPE column and plugin marketplace suffix dropped.
show --plain: parsable "type<TAB>name" list, no status, zero claude calls —
the contract for the upcoming design gate. All derived from .profile files.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X3e8LaH2vymmxyh36h3jFU
PASS A done-detection was ~60% machinery a capable agent already does for
free (the baseline checked done tasks and left the umbrella task alone
unaided). Cut the git-command how-to and worked example; keep only the
load-bearing restraint rule (flip only on a clean task<->commit map;
partial/umbrella/vague stay unchecked; never guess).
Add a Red flag: the STEP 3 gate STOP was never exercised (non-interactive
build harness printed the gate then proceeded as approved) — confirm it
halts before any write on first real use. TDD note records the same caveat.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X3e8LaH2vymmxyh36h3jFU
Two-mode capitalize: default pre-wipe flush, --ritual adds the 3-question
end-of-session reflection (now deduped, unlike legacy /close). New STEP 2B
reconciles .claude/tasks/TODO.md — PASS A done-detection (only on an
unambiguous task<->commit map), PASS B explicit-only capture with an
anti-noise filter (never track commit/deploy/push/release/tag) and BDR
routing for orientation directives. STEP 3 gate gains a separate TODO block;
journal/handoff report TODO ops. /close becomes a thin alias for
/capitalize --ritual (zero duplicated logic).
Built via superpowers:writing-skills TDD: RED baseline (no skill) folded a
push/tag parasite into the TODO, invented a subtask, and wrote with no gate;
GREEN re-run on the same fixture stops at the gate, drops both dups (footer
shows existing IDs), logs one learning, checks only the cleanly-done task,
ignores the parasite, and routes the GraphQL directive to BDR.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X3e8LaH2vymmxyh36h3jFU
readme-updater.md does not exist on disk; STEPs 5b and 12 were broken
references. doc-syncer.md AUTO MODE covers both cases (README bootstrap
when missing, drift sync after changes).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Round 2 of darwin optimization (judge-identified residuals):
- 3c marker rule now cross-references the STEP 0 dangling/corrupted
exceptions instead of contradicting them
- corrupted-but-present state JSON branch defined (trust no axis,
ask repair/reset; headless -> full codebase report-only, file as-is)
- unreachable user at 3e max-cycles STOP -> fail closed: revert axis
fixes, findings back to open, marker untouched
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Round 1 of darwin optimization, dim3 (failure-mode encoding). Live test
showed two agents diverging on undefined branches:
- dangling marker + unreachable user -> now full-codebase report-only,
marker untouched (corrupted state needs user-approved repair)
- no axes named + unreachable user -> now defaults to all four axes
Also adds the matching Common-mistakes row. Includes test-prompts.json.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Strip the disable-model-invocation frontmatter key from all 19 editable skills. Absent = default = model invocation enabled. 8 were 'true' and blocked the model AND orchestrators from self-routing (status, plugin-check, analyze, onboard, refactor, init-project, pdf-translate, ship-feature) — contradicting the CLAUDE.md skill-routing rules. The other 11 were 'false', a no-op noise line.
The setting is binary (no per-caller granularity), so enabling orchestrator chaining also enables model auto-fire — accepted. Genuinely destructive operations remain guarded by the careful/guard hooks, independent of this flag.
Capitalized: BDR-019 (decision), LRN-026 (learning), journal 2026-06-09.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Salvages registry-worthy insights from the conversation before /clear or /compact wipes context. Scans the conversation, dedups each candidate against the existing .claude/memory/ registries (its signature move vs /close), routes across all 5 registries behind a compact approval gate, always writes a journal line.
Baseline-tested per superpowers:writing-skills (RED/GREEN/REFACTOR): the no-skill baseline double-logged one incident across LRN+BLK; the skill passes clean and now counters that via a 'one incident -> one primary registry' rule. Ships v1 with the approval gate as the human safety net (same posture as /prune-memory).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Centralize gstack toggling in the `profile` command without losing the
active-profile label.
- `gstack on` re-enables ALL parked gstack skills (moves
skills-disabled/gstack__* back) but does NOT touch .active-profile,
so the user layers full gstack on top of their current profile and
the statusline label is preserved. Unlike `reset`, which clears the
label to "none".
- `gstack off` disables gstack skills not listed in the active profile;
errors cleanly when no profile is active (needs one to know what to
keep).
Refactor (behavior-preserving): extract three shared helpers
`enable_all_gstack`, `disable_gstack_not_in`, `parked_gstack_count` and
rewire `cmd_reset` + `cmd_set` to reuse them instead of duplicating the
symlink-toggle loops. Wire `gstack` into main() dispatch, usage(), and the
header usage block.
Docs: SKILL.md argument-hint, examples, and output-policy updated. The
generic `make profile cmd="gstack on"` target already covers Make usage.
Verified: shellcheck CLEAN, `bash -n` OK, 6-case test (help, bad-action,
off-with-no-profile, on, off-trim, on-cycle) with final assertion that the
live symlink state was restored exactly to its pre-test value.
Memory: capitalize BDR-018 (decision), LRN-024 (DRY helper-extraction
pattern), BLK-007 (6 gstack source skills ios-*/spec unlinked post
submodule bump — open follow-up), EVAL-002 (self-eval, false "full.profile
bug" flag corrected pre-edit). Backfill index drift: BDR-017, BLK-005/006.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
OCR/image-based PDF pipeline: convert pages to PNGs, read with Claude
Vision (bypasses unreliable OCR text layer), translate with cross-page
glossary consistency, reconstruct faithful HTML via /design-html.
5 steps: deps check → page images + assets → style analysis →
page-by-page read+translate → HTML reconstruction → visual QA.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Major SKILL.md rewrite:
- Fast path: skip extraction when graph exists and user asks a question
- Gemini backend replaces Kimi as default external LLM
- All file I/O uses ensure_ascii=False + encoding="utf-8"
- Monorepo support via per-subfolder extraction + merge
- Obsidian/HTML export via CLI instead of inline Python
- Node ID format includes parent dir to prevent ghost duplicates
- file_type gains "concept" as valid value
- Subagent chunk paths must be absolute
- --help flag prints usage and stops
- Large corpus gate raised from 200 to 500 files
Bumps gstack submodule to 026751e.
Co-Authored-By: Claude <noreply@anthropic.com>
Superset of web-full + dev + audit + plan tools — covers brainstorm →
design → architecture review → scaffold → implement → ship → audit
pipeline needed by /init-project to produce a real MVP in one session.
Also renames cmd_current "no-profile" sentinel from "full" to "none"
to avoid collision with the new profile name.
Co-Authored-By: Claude <noreply@anthropic.com>
Skills covered: close, graphify, harden, profile, prune-memory.
Used by /darwin-skill dim 8 effect testing.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
First end-to-end run of /prune-memory on real .claude/memory/ surfaced
a broken verify script:
Old: `prefix=$(basename "$f" .md | tr a-z A-Z | cut -c1-3)` derived
the prefix from the filename's first 3 letters → produced DEC / LEA /
BLO. Actual prefixes are BDR / LRN / BLK. The grep then matched zero
entries, no MISSING/ORPHAN was ever reported, and the script printed
its "OK if blank" footer regardless of real state. False clean signal.
Fixed: hard-mapped filename → prefix via `declare -A PREFIX_MAP`.
Verified against current registries — 14 BDR + 16 LRN + 2 BLK + 1 EVAL
entries all index-consistent, no false negatives.
Added EVAL prefix to the map (evals.md was missing from the loop in
v1). Footer line clarified to `(blank above = OK)`. `wc -l` excludes
`.original.md` backups from the output.
Note: caveat in skill body said "v1 ships without baseline TDD test —
STEP 2 approval gate is the safety net". First real test caught a
verify bug that bypassed STEP 2 entirely. Lesson: STEP 4 is its own
safety net and needs its own test.
Co-Authored-By: Claude <noreply@anthropic.com>
Two issues fixed across personal skills:
1. Five skills had frontmatter > 1024 chars (spec limit per agentskills.io):
- client-handover: 1920 → 924
- doc: 1390 → 734
- seo: 1378 → 885
- geo: 1189 → 734
- validate: 1050 → 742
All now compliant.
2. Three orchestrators had workflow-summary descriptions that create
"shortcut" risk per writing-skills CSO (Claude may follow the
description instead of reading the full skill body, skipping steps):
- ship-feature: "design → plan → implement (TDD) → review → finish"
→ "Use when shipping a feature end-to-end — needs… (9-step pipeline)"
- init-project: "interview → design → scaffold → implement (TDD)"
→ "Use when initializing a brand-new project from scratch — needs…"
- onboard: "detect archetype, install claude-config, run full audit…"
→ "Use when bringing an existing repo into the framework — needs…"
Descriptions now follow "Use when [triggers]…" pattern. Workflow detail
preserved in skill body where it belongs. Triggers consolidated and
trimmed of duplicates.
Co-Authored-By: Claude <noreply@anthropic.com>
Old rule `li input[type="checkbox"] + *` absolutely-positioned the first
element sibling after the checkbox (typically <a>, <code>, <strong>),
yanking links and code spans out of flow and overlapping adjacent
content in the rendered PDF.
Replace with a targeted rule that styles the native disabled checkbox
inline (small green box) and leaves siblings untouched. Pandoc GFM emits
`<li><input disabled type="checkbox"> text…</li>` with no wrapper class,
so we target `li > input[type="checkbox"]` directly.
Co-Authored-By: Claude <noreply@anthropic.com>
NAP table now includes a "Description courte" / "Short description" field
(1-2 sentences, lifted from hero/meta description, pasted identically
across Google Business, Bing Places, Apple Maps, directories). Same field
client was already typing 10 different ways across platforms — now it's
the single source of truth.
BrightLocal Free Tools page was retired in 2026 (service now paid-only).
Replaced with Moz Local Citation Checker (free 60s audit across 50+
directories, no credit card). Both FR and EN versions updated.
Co-Authored-By: Claude <noreply@anthropic.com>
Restructure deliverable from 4 to 6 chapters:
- §2 (new): score table promoted from technical annex to top of doc for
immediate visual proof of impact (tested with local-business clients —
converts "what did I pay for?" doubt within 30 seconds).
- §4 (new): NAP table promoted from §7 annex so client reads identity
values (name, address, phone, hours, categories, short description)
BEFORE attacking §5 todo list. Prevents 10-different-description drift
across external platforms that degrades Google's NAP-consistency signal.
- §5 (todo) and §6 (tech details) renumbered; §7/§8 annexes still optional.
Pandoc bumped to gfm+gfm_auto_identifiers so internal anchor links like
[§4](nap) resolve in the rendered HTML/PDF.
Co-Authored-By: Claude <noreply@anthropic.com>
Two distinct render bugs producing overlapping text on multi-page PDFs:
1. Bare-URL duplication. The print stylesheet injects `(href)` after every
external link via `a[href^="http"]::after`. When pandoc/marked
auto-links a bare URL or renders `[X](X)`, the visible text already
equals the href, so the pseudo-element produces "URL (URL)" and the
trailing duplicate wraps onto the next line, colliding with the
following block (e.g. "https://pagespeed.web.dev/ (https://...)" then
"• Ouvrir, taper l'URL...").
Fix: post-process the body HTML in handover-to-pdf.sh; tag every
`<a href="X">X</a>` (text == href, ignoring trailing slash + case)
with `class="bare-url"`, and exclude `a.bare-url::after` from the
URL-injection rule. Named links still get `(URL)` for print legibility.
Belt-and-braces: add `white-space: nowrap` and `break-inside: avoid`
on the remaining `::after` so future long URLs cannot wrap across
page boundaries either.
2. List item splitting across page boundary. `li` had only
`orphans/widows: 3` and no `break-inside`, so a long item could put
its bullet on page N and its text on page N+1, overlapping unrelated
content. Heading-to-first-block adjacency was also unprotected, so
"heading at bottom of page A / intro paragraph or first bullet at top
of page B" could produce visual overlap during reflow.
Fix: add `li { page-break-inside: avoid; break-inside: avoid; }` and
`h{1..4} + p|ul|ol { break-before: avoid; }` so list items stay
intact and intros stay glued to their heading.
Verified end-to-end: rendered sample md with bare URL + named link +
heading-followed-by-list straddling a page break; pdftotext shows each
URL once, no orphaned bullets, no `::after` warning from weasyprint.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three bugs surfaced on the LIVRAISON.pdf test render:
1. **MD→HTML conversion** — host had no pandoc, no python-markdown, fell
back to `npx marked < "$src"`. marked CLI 16.x ignores stdin and
dumps its own cli.js source. Resulting PDF body = marked's binary
source (`#!/usr/bin/env node`, `Marked CLI`, copyright). Fix:
`npx --yes marked --gfm -i "$src"` (file path via -i, not stdin).
2. **Cover background** — original cream `#F5F0EB` + 8mm green stripe
was washed out. Iterated to white-pure bg with subtle radial
sage/forest tints, black-deep title, green-forest accents
(eyebrow, meta labels, footer, border). Solid green-dark tried
first then rejected (too heavy for long client-facing doc).
3. **Default logo** — SVG `logo-horizontal.svg` rendered cream-toned,
blended into bg. Switched LOGO_URL default to
`https://zenquality.fr/assets/logo-horizontal-1024.png`.
Also added test-artifact gitignore rules for LIVRAISON.* / HANDOVER.*
project-local renders.
Verified: regenerated LIVRAISON.pdf → 164 KB, 19 pages, full content
rendered, white cover with black title + green-forest accents +
visible PNG logo.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
End-to-end pipeline that hardens the project, commits, pauses for
deploy, validates the live site, then generates a non-technical client
deliverable (LIVRAISON.md / HANDOVER.md):
1. /seo (SEO+GEO) and /harden run in parallel with auto-fix loops
until each scores >=17/20.
2. /commit-change + push if changes were made.
3. Pause to tell the user what to deploy and wait for confirmation.
4. /validate against the live site.
5. Per-audit gate >=17/20 — stop and analyze if any below.
6. Write client doc with before/after score table + owner-maintenance
checklist.
Reads git history + .claude/memory/ registries for context. For
local-business projects, appends a manual NAP-consistency platform
checklist (Google Business, Pages Jaunes, Yelp, Facebook, Instagram,
TikTok, Apple Maps, Bing Places, TripAdvisor). Optional build/deploy
chapter on prompt.
Adds skills/client-handover/SKILL.md (slash-command entrypoint),
skills/client-handover/checklists/seo-geo-manual.md (NAP checklist),
agents/client-handover-writer.md (orchestrator agent).
Co-Authored-By: Claude <noreply@anthropic.com>
Auto-discover what the project actually has instead of a fixed doc list:
root files (incl. DEPLOY.md, SECURITY.md, ARCHITECTURE.md, ROADMAP.md),
docs/**, and .claude/{tasks,audits,memory}/.
Detect stack (Node/Python/Rust/Go/Ruby/PHP/Dart/.NET) and deploy
complexity (NONE/TRIVIAL/NON_TRIVIAL) to drive doc-needs:
- Propose DEPLOY.md only when non-trivial (Docker, fly.toml, k8s,
multi-stage CI).
- Propose inlining/removing DEPLOY.md when deploy is trivial.
- Enforce README presence with typical GitHub layout.
Add CREATE/REMOVE proposal categories to the validation gate. Update
auto-mode to map deploy artifacts to DEPLOY.md and decisions.md
architectural changes back to CLAUDE.md/README.
Sync skills/doc/SKILL.md description + triggers to match.
Co-Authored-By: Claude <noreply@anthropic.com>
- GitHub URL clone support (single + multi-repo merge)
- uv tool detection ahead of pipx shebang
- Optional Kimi K2.6 backend via MOONSHOT_API_KEY
- Discrete confidence rubric (0.55/0.65/0.75/0.85/0.95) replaces continuous range
- Aggregated community view for graphs >5000 nodes
- Manifest persistence after --update so next run diffs against current state
- Video file support, rationale-as-attribute (not separate node)
Co-Authored-By: Claude <noreply@anthropic.com>
Ship lib/profile.sh + 9 profiles in lib/profiles/. A profile is a
plain-text file listing items + types (gstack | personal | external |
plugin@<marketplace> | mcp | cli). `profile set <name>` enables the
listed items and disables the rest:
- gstack/personal/external skills: symlink toggle skills/ ↔
skills-disabled/ (gstack__<name> prefix to avoid collisions; no
prefix for personal/external).
- plugins typed `plugin@<marketplace>`: actually toggled via
`claude plugin enable|disable <name>@<marketplace>`. Allowlist:
MANAGED_PLUGINS = ui-ux-pro-max, plugin-dev, pr-review-toolkit.
Denylist: PROTECTED_PLUGINS = caveman, security-guidance,
superpowers (always-on, never disabled even if absent from a
profile).
- mcp magic: delegated to lib/toggle-external.sh which already
handles the MAGIC_API_KEY env lookup. Other MCPs stay advisory.
- cli (rtk, gsd, ctx7, graphify): status-only, never auto-installed.
Profiles shipped:
web public website work — frontend + content + light dev
seo SEO + GEO + W3C audit (search/AI indexability + a11y)
web-full production website end-to-end (web ∪ seo ∪ qa-only/canary)
backend backend / API / system dev — no design, no SEO
design visual QA, design systems, mockups, polish
dev daily code work — features, fixes, refactor, ship
qa site testing, perf, canary, validation
audit comprehensive audit — security + SEO + perf + health
minimal strip all gstack skills (quiet session)
Commands:
profile list / show <name> / current / apply <name> / set <name> /
reset / diff <a> <b>
`current` heuristic returns "full" when nothing is disabled, otherwise
picks the profile with the highest available-ratio (counts both
"enabled" and "installed" — the latter for CLIs). Tiebreaker: larger
profile total wins, so web-full beats web at a 100% tie.
`reset` re-enables every gstack skill but does NOT touch plugins —
the user re-enables a managed plugin manually or via `apply <profile>`.
This is documented in the trailing info line.
Integration:
- skills/profile/SKILL.md — `/profile` slash command, lists profiles,
documents the per-type mechanism, points at lib/profile.sh.
- agents/plugin-advisor.md — DETECT phase calls `profile current`,
OUTPUT adds a PROFILE line, and TOGGLING EXTERNAL TOOLS gains a
"Skill profiles" section with a signal → profile recommendation
table.
- lib/toggle-external.sh — header pointer to profile.sh for fine-
grained activation (toggle-external still owns whole-gstack and
magic-MCP toggles).
- Makefile — `make profile cmd="set <name>"`, profile-list,
profile-current, profile-reset.
Tested end-to-end: `set web` enables ui-ux-pro-max + magic; `set seo`
disables ui-ux-pro-max; `set minimal` disables ui-ux-pro-max but
spares always-on plugins; `reset` restores all 64 skills; shellcheck
clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Registries are re-read at every session start; consistency pays back
each time. Adding a single rule in CLAUDE.md § Memory registries
(authoritative) + a short reminder at the end of each CAPITALIZE
block (ship-feature, bugfix, hotfix, feat, commit-change, close).
Rationale: (1) model re-reads the registries more efficiently in a
single language, (2) lower token cost for English (model's primary
training language), (3) easier cross-project reuse. Interactive gates
may still mirror the user's language — only the written entry is
constrained.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Before: SEO.md, GEO.md, HARDEN.md, VALIDATE.md, BUGS-FOUND.md landed
at project root. After: all five go to .claude/audits/. Covers both
dispatcher write paths and the dispatcher bash commands that parse
each report (test -s, grep score, wc) — otherwise the dispatcher
would look for the file at the old location.
- skills/seo,geo,harden,validate,code-clean — write paths + console
summaries ("Report: .claude/audits/X.md")
- skills/harden,validate — bash parsing commands (test/grep/wc) aligned
- agents/seo-analyzer,validator-analyzer,code-cleaner — agent-side refs
- agents/validator-analyzer frontmatter description updated
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Manual fallback for the 3-question ritual (what was decided / learned
/ blocked) when none of the completion skills ran. Pre-fills BDR/LRN/
BLK candidates from git log + conversation context, presents for
validation, appends approved entries to .claude/memory/ registries
plus a timeline line in journal.md. Always writes the journal line
even if all 3 questions are skipped.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Registries only get filled if something actively writes to them.
Without integration, the 3-question ritual is aspirational text.
Adds a CAPITALIZE step at the end of every completion skill so work
milestones automatically propose BDR/LRN/BLK entries from context.
- ship-feature STEP 9 — decision/learning/blocker candidates per feature
- bugfix STEP 7 — always propose BLK with root cause; LRN if pattern reusable
- hotfix STEP 5 — default skip; only prompt when non-obvious lesson surfaces
- feat STEP 6 — propose BDR for design choice, LRN for pattern
- commit-change Phase 4 — analyze the commit batch, propose grouped entries
Every variant also appends a one-liner to .claude/memory/journal.md under
today's date heading.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adapt the two skills that scaffold project config + the underlying
onboarder agent + the dotfiles-meta archetype folder tree to the new
governance layout. AUDIT_* files go to .claude/audits/, TODO goes to
.claude/tasks/, and .claude/memory/ is seeded from ~/.claude/templates/memory/
with all 5 registries (decisions/learnings/blockers/journal/evals).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
