Menu
One-time project setup — creates directories, config, and initial files.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Basé sur la classification professionnelle SOC
Re-organize recall knowledge/workflow topics into their best shape — unify topic filenames, frontmatter, and titles, merge overlapping topics, compact stale content, extract shared general principles into their own topics, and rewrite the index coherently. Use whenever the user says "re-org the recall topics", "clean up recall", "make the recall coherent", "extract a general topic", "tidy the knowledge index", or when recall files have drifted (mismatched names, duplicated content, oversized files, a stale index). Keeps recall well-managed as a whole rather than fixing one file at a time.
Overview of all tracked branches (active, stale, merged, archived).
Promote branch knowledge to project level after merge.
Show recall state — branch, task, loaded topics, stale/unpromoted info.
Add knowledge entry with scope control (--project/--branch flags).
Search across all knowledge files including archives.
| name | recall-init |
| description | One-time project setup — creates directories, config, and initial files. |
Set up the recall system for the current project.
recall.root or default ~/.local/share/claude/recall/).git remote get-url origin.<storage-root>/<project>/directives.md — project config and rules (no markdown headings, no YAML fences):
auto-save.auto: [hardware findings, build errors, debugging root causes, API behaviors]
auto-save.ask: [coding conventions, architecture decisions]
auto-save.never: [temporary workarounds, debugging session logs]
auto-save.default: auto
promotion: auto
stale-branch-days: 30
default-branch: develop
confidence-min: observed
maintenance:
status-max-lines: 150
status-final-lines: 100
topic-max-lines: 200
task-knowledge-split-lines: 150
Add project-specific rules as bullet points after the config block. Only add rules the user requests — no placeholders.
knowledge/index.md — empty file (topics added as discovered) workflows/index.md — empty file (workflows added as discovered) user.md — key-value pairs only, no headings or boilerplate:
CONTAINER=<ask user>
WORKSPACE=<ask user>
editor: <ask user or omit>
commit-style: <ask user or omit>
test-prefs: <ask user or omit>
Only include fields the user provides. Omit any field they don't specify.
${CLAUDE_PLUGIN_ROOT}/scripts/create-branch-dir.sh --project-dir <project-dir> --branch <branch-name> --parent <parent-branch> --mode <full|lightweight>
When writing to status.md or knowledge files, check line counts against maintenance limits from directives.md:
<!-- maintenance: compact needed --> at line 1. Do not compact now.### subsections into sibling files, update index.knowledge/ directory with per-topic files + index.md.Hard compaction (rewriting status.md to compact format) happens only at lifecycle events (/task-complete, /task-abandon) or when session-start detects an oversized file or the compact-needed marker.
Every knowledge file read is also a quality check. Fix in-place (Tier 1): factual errors, stale references, duplicates, vague claims — only when you have the precise replacement in current context. Maximum 2 quality edits per file read. Never modify benchmark tables, commit hashes, code snippets, or quoted output. Flag cross-file issues as <!-- quality-review: ... --> comments (Tier 2) for later resolution.