read-continue-here
À propos
Cette Compétence Claude reprend automatiquement le travail à partir de sessions interrompues en lisant un fichier de transmission CONTINUE_HERE.md au début de chaque session. Elle gère la détection du fichier, la vérification de sa fraîcheur, l'analyse structurée et la confirmation de l'utilisateur avant le nettoyage. Utilisez-la pour redémarrer après des interruptions ou pour mettre en place des flux de travail de continuation automatique.
Installation rapide
Claude Code
Recommandénpx skills add pjt222/agent-almanac -a claude-code/plugin add https://github.com/pjt222/agent-almanacgit clone https://github.com/pjt222/agent-almanac.git ~/.claude/skills/read-continue-hereCopiez et collez cette commande dans Claude Code pour installer cette compétence
Documentation
Read Continue Here
Read structured handoff file. Resume work from where prior session stopped.
When Use
- Start new session. CONTINUE_HERE.md sits in project root
- After SessionStart hook injects handoff context
- Bootstrap identity. Detect prior session artifacts
- Set up auto continuation detection for project (one-time infra)
Inputs
- Required: Project directory (default: cwd)
- Optional: Configure infra (SessionStart hook + CLAUDE.md instruction)
- Optional: Delete file after consume (default: yes)
Steps
Step 1: Detect and Read Handoff File
Check for CONTINUE_HERE.md in project root.
ls -la CONTINUE_HERE.md 2>/dev/null
Absent? Exit gracefully — nothing to continue from.
Present? Read contents. Parse 5 sections: Objective, Completed, In Progress, Next Steps, Context. Pull timestamp and branch from header line.
Got: File read. Sections parsed into clear mental model of prior session state.
If fail: File exists but malformed (missing sections, empty)? Treat as partial signal — pull what is there. Note gaps to user.
Step 2: Check Freshness
Compare file timestamp against current time.
# File modification time
stat -c '%Y' CONTINUE_HERE.md 2>/dev/null || stat -f '%m' CONTINUE_HERE.md
# Current time
date +%s
Classify freshness:
- Fresh (< 24 hours, same branch): safe to act on direct
- Stale (> 24 hours or different branch): flag to user before proceed
- Superseded (new commits exist after handoff timestamp): someone worked on project since handoff
Check branch alignment.
git branch --show-current
git log --oneline --since="$(stat -c '%Y' CONTINUE_HERE.md | xargs -I{} date -d @{} --iso-8601=seconds)" 2>/dev/null
Got: Freshness verdict (fresh, stale, superseded) with evidence.
If fail: Not in git repo? Skip branch and commit checks. Use timestamp in file header alone.
Step 3: Summarize and Confirm Resume
Show handoff state to user concise:
- "Prior session objective: [Objective]"
- "Completed: [summary]"
- "In progress: [summary]"
- "Proposed next action: [Next Steps item 1]"
Stale or superseded? Show evidence. Ask whether proceed with handoff or start fresh.
Next Steps items tagged **[USER]**? Surface them explicit — they need user decision before work proceed.
Got: User confirms resume plan, maybe with tweaks. Agent has clear mandate for next action.
If fail: User says "start fresh" or "ignore that file"? Acknowledge. Proceed without handoff context. Offer delete file to prevent future confusion.
Step 4: Act on Handoff
Start work from Next Steps item 1 (or where user pointed):
- Reference In Progress items to grasp partial state
- Use Context section to avoid retry of failed approaches
- Treat Completed items as done — no re-verify unless user asks
Got: Agent productive on right task, informed by handoff file.
If fail: Next Steps ambiguous or In Progress unclear? Ask user for clarity. No guess.
Step 5: Clean Up
Handoff consumed and work underway? Delete CONTINUE_HERE.md.
rm CONTINUE_HERE.md
Stale handoff files cause confusion in future sessions.
Got: File removed. Project root clean.
If fail: User wants keep file (e.g., as reference during session)? Leave it but note: must delete before session end. Else next session re-consumes.
Step 6: Configure SessionStart Hook (Optional)
Not yet configured? Set up auto reading of CONTINUE_HERE.md on session start.
Create hook script:
mkdir -p ~/.claude/hooks/continue-here
cat > ~/.claude/hooks/continue-here/read-continuation.sh << 'SCRIPT'
#!/bin/bash
# SessionStart hook: inject CONTINUE_HERE.md into session context
# OS-aware: works on native Linux, WSL, macOS, and Windows (Git Bash/MSYS)
set -uo pipefail
# --- Platform detection ---
detect_platform() {
case "$(uname -s)" in
Darwin) echo "mac" ;;
Linux)
if grep -qi microsoft /proc/version 2>/dev/null; then
echo "wsl"
else
echo "linux"
fi ;;
MINGW*|MSYS*|CYGWIN*) echo "windows" ;;
*) echo "unknown" ;;
esac
}
PLATFORM=${PLATFORM:-$(detect_platform)}
CONTINUE_FILE="$PWD/CONTINUE_HERE.md"
if [ ! -f "$CONTINUE_FILE" ]; then
exit 0
fi
# Strip CRLF (files on NTFS often have Windows line endings)
CONTENT=$(sed 's/\r$//' "$CONTINUE_FILE")
# JSON-escape: prefer jq, fall back to portable awk
if command -v jq >/dev/null 2>&1; then
ESCAPED=$(printf '%s' "$CONTENT" | jq -Rsa .)
else
ESCAPED=$(printf '%s' "$CONTENT" | awk '
BEGIN { ORS=""; print "\"" }
{
gsub(/\\/, "\\\\")
gsub(/"/, "\\\"")
gsub(/\t/, "\\t")
if (NR > 1) print "\\n"
print
}
END { print "\"" }
')
fi
cat << EOF
{"hookSpecificOutput":{"sessionStartContext":{"additionalContext":$ESCAPED}}}
EOF
SCRIPT
chmod +x ~/.claude/hooks/continue-here/read-continuation.sh
Add to ~/.claude/settings.json in SessionStart hooks array.
{
"type": "command",
"command": "~/.claude/hooks/continue-here/read-continuation.sh",
"timeout": 5
}
Got: Hook script exists, executable, registered in settings.json. Next session start: if CONTINUE_HERE.md exists, content injects into session context.
If fail: Verify settings.json valid JSON after edit. Test hook manual: cd /your/project && ~/.claude/hooks/continue-here/read-continuation.sh. Script falls back to awk if jq not installed — jq recommended, not required.
Step 7: Add CLAUDE.md Instruction (Optional)
Add brief instruction to project CLAUDE.md so Claude knows file purpose.
## Session Continuity
If `CONTINUE_HERE.md` exists in the project root, read it at session start. It contains a structured handoff from a prior session: objective, completed work, in-progress state, next steps, and context. Act on it — acknowledge the continuation, summarize prior state, and propose resuming from the Next Steps section. If the file is older than 24 hours, flag this to the user before proceeding. After the handoff is consumed, the file can be deleted.
Got: CLAUDE.md has instruction. Future sessions read and act on CONTINUE_HERE.md even when SessionStart hook not configured.
If fail: CLAUDE.md absent? Create with this section only. File too long? Add instruction near top where it not get truncated.
Checks
- CONTINUE_HERE.md detected (or absence handled gracefully)
- Freshness checked (timestamp, branch, post-handoff commits)
- Resume plan shown to user, confirmed
- Work began from correct Next Steps item
- File cleaned up after consume
- (Optional) SessionStart hook script exists, executable
- (Optional) CLAUDE.md has session continuity instruction
Pitfalls
- Act without confirm: Always show resume plan to user. They may have changed mind on what to work on, even when file fresh.
- Trust stale files blind: Handoff file older than 24 hours or from different branch = suggestion, not mandate. Always check freshness.
- Ignore Context section: Most valuable part = failed approaches. Skip it = retry dead ends.
- Forget cleanup: Leave CONTINUE_HERE.md after consume = next session re-acts on it = confusion.
- Treat Completed as unverified: Unless user asks specific, no re-do completed work. Trust prior session.
See Also
write-continue-here— complement: write handoff file at session endbootstrap-agent-identity— full identity reconstruction; uses handoff detect as one heuristicmanage-memory— durable cross-session knowledge (complement to this ephemeral handoff)write-claude-md— project instructions where optional continuity guidance lives
Dépôt GitHub
Compétences associées
executing-plans
DesignUtilisez la compétence executing-plans lorsque vous disposez d'un plan de mise en œuvre complet à exécuter par lots contrôlés avec des points de contrôle de revue. Elle charge et examine le plan de manière critique, puis exécute les tâches par petits lots (3 tâches par défaut) tout en rapportant la progression entre chaque lot pour une revue par l'architecte. Cela garantit une mise en œuvre systématique avec des points de contrôle de qualité intégrés.
requesting-code-review
DesignCette compétence délègue un sous-agent réviseur de code pour analyser les modifications apportées au code par rapport aux exigences avant de poursuivre. Elle doit être utilisée après avoir terminé des tâches, implémenté des fonctionnalités majeures, ou avant une fusion vers la branche principale. La revue aide à détecter précocement les problèmes en comparant l'implémentation actuelle avec le plan initial.
connect-mcp-server
DesignCette compétence fournit un guide complet permettant aux développeurs de connecter des serveurs MCP à Claude Code via les transports HTTP, stdio ou SSE. Elle couvre l'installation, la configuration, l'authentification et la sécurité pour intégrer des services externes tels que GitHub, Notion et des API personnalisées. Utilisez-la lors de la configuration d'intégrations MCP, de la configuration d'outils externes ou du travail avec le Protocole de Contexte de Modèle de Claude.
web-cli-teleport
DesignCette compétence aide les développeurs à choisir entre les interfaces Web et CLI de Claude Code en fonction de l'analyse des tâches, puis permet une téléportation transparente des sessions entre ces environnements. Elle optimise le flux de travail en gérant l'état et le contexte de la session lors du passage entre le web, la CLI ou le mobile. Utilisez-la pour des projets complexes nécessitant différents outils à diverses étapes.
