wf-resume

name: wf-resume description: Resumes an interrupted waterfall workflow — preflight, need lookup, Agent Teams team re-creation, OR handoff with resume brief. user-invocable: false allowed-tools: Read, Grep, Glob, Bash, AskUserQuestion, Skill, TeamCreate

wf-resume — Resume an interrupted workflow

This skill is the entry point to resume a paused or interrupted waterfall need. It is invoked by the /waterfall:resume command.

Resume flow

Step 1 — Preflight (fail-fast)

bash ${CLAUDE_PLUGIN_ROOT}/scripts/wf-check-bash.sh

If not running under bash (Git Bash on Windows) → the script fails with an explicit message. STOP.

bash ${CLAUDE_PLUGIN_ROOT}/scripts/wf-check-teams.sh

If the CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS flag is missing → the script fails with an explicit message. STOP.

jq verification (used by all wf scripts to parse .wf-state.json, ack-registry.json, etc.):

INSTALL_CMD=$(bash ${CLAUDE_PLUGIN_ROOT}/scripts/wf-check-jq.sh) || JQ_RC=$?

• Exit 0 → continue.
• Exit 2 → jq missing. If $INSTALL_CMD non-empty, AskUserQuestion: "jq is required. Install it via ${INSTALL_CMD}?" (Yes / No). On Yes: bash -c "$INSTALL_CMD" then re-run the check. On No or if command empty: display stderr and stop.

Step 1.a — Resolve WF_SID (new HO window)

Critical for statusline (INV-002): on resume, $CLAUDE_SESSION_ID is a fresh sid (new Claude Code window) that does not match the session_id stored in .wf-state.json. We must force WF_SID to the current HO sid so the --reactivate migration takes place.

export WF_SID="${CLAUDE_SESSION_ID}"

Without this, the statusline (which filters on state.session_id == $CLAUDE_SESSION_ID) will not display the resumed need.

Step 1.bis — Same-sid ambiguity detection (EX-005)

Note: when resuming in a new window, WF_SID is a fresh sid that won't match any state. This step is effectively a no-op on resume; it remains useful if /waterfall:resume is invoked in the same window as a workflow that was just abandoned.

If WF_SID is set and no <name> is provided as argument:

# Count needs whose session_id == $WF_SID
matches=()
for state in C:/projets/waterfall/wf/needs/*/.wf-state.json; do
    [[ -f "$state" ]] || continue
    state_sid=$(jq -r '.session_id // ""' "$state" 2>/dev/null)
    [[ "$state_sid" == "$WF_SID" ]] && matches+=("$(dirname "$state" | xargs basename)")
done
if [[ ${#matches[@]} -ge 2 ]]; then
    # AskUserQuestion: pick among the needs (one option per need)
fi

Step 1.quater — Config read

source ${CLAUDE_PLUGIN_ROOT}/scripts/wf-read-config.sh
# No template copy (need directory already exists)
# watchdog and agent_mode apply normally

Conditional watchdog (after OR re-spawn)

if [[ "$WF_WATCHDOG_INTERVAL" != "off" ]]; then
  DELAY_MIN="${WF_WATCHDOG_INTERVAL//min/}"  # "3min" → "3"
  CronCreate(cron: "*/${DELAY_MIN} * * * *", prompt: "watchdog tick wf-<name>", recurring: true)
  # Belt-and-suspenders: touch the marker that OR checks
  echo "<cron_job_id>" > wf/needs/<name>/.watchdog-cron-active
  # HO message: "Watchdog active (${WF_WATCHDOG_INTERVAL})"
fi

Step 2 — Need lookup via wf-orchestrate.sh

bash ${CLAUDE_PLUGIN_ROOT}/scripts/wf-orchestrate.sh --list

Returns a JSON array of needs with {name, phase, step, status, last_updated}.

Cases:

• No active need: inform HO, STOP
• Single active need: automatic selection as <name>
• Multiple active needs: AskUserQuestion to HO to pick (one option per need, label = name, description = <phase>:<step> [status])

Step 3 — Orphan worktree detection (optional MVP)

git worktree list --porcelain

Identify worktree-dv* entries whose directory no longer exists. If orphans detected, propose cleanup to HO via AskUserQuestion. If approved: git worktree prune.

Step 4 — Lookup table phase → agents to re-spawn

Read bash ${CLAUDE_PLUGIN_ROOT}/scripts/wf-orchestrate.sh <name> --query to know the current_phase.

Step 5 — Team re-creation and OR spawn

1. Load the wf-pm skill via Skill({name: "wf-pm"}). The main conversation thus adopts PM responsibilities — PM is never spawned as a separate agent (aligned with wf-new step 3). The team-lead created by TeamCreate IS the main.
2. PM (= main) executes TeamCreate according to agent_mode:

   if [[ "$WF_AGENT_MODE" == "team" ]]; then
     TeamCreate wf-<name>
   else
     # Subagent mode: spawn OR via Agent tool, no TeamCreate
   fi
   

3. PM (= main) may clear the traceability registry (optional, DEC-001):

   bash ${CLAUDE_PLUGIN_ROOT}/scripts/wf-registry.sh clear <name>
   

   DEC-001: traceability only — enforcement uses agent_type from the harness payload. Skipping this step does not prevent teammates from completing their steps.

4. PM (= main) spawns OR first via Agent (subagent_type=waterfall:wf-or, name=or, team_name=wf-<name>). For traceability (optional):

   bash ${CLAUDE_PLUGIN_ROOT}/scripts/wf-registry.sh add <name> <or_agent_id> or
   

5. PM (= main) sends the resume brief to OR via SendMessage:

<brief>
  <task_id>BRIEF-RESUME-<timestamp></task_id>
  <phase>RESUME</phase>
  <besoin><name></besoin>
  <besoin_dir>wf/needs/<name>/</besoin_dir>
  <action>resume</action>
  <current_phase><phase read from --query></current_phase>
  <context>
    <note>Session interrupted. Read .sdd-state.json via --query and or.log to reconstruct the state.</note>
  </context>
</brief>

4. OR executes its resume sequence: a. bash ${CLAUDE_PLUGIN_ROOT}/scripts/wf-orchestrate.sh <name> --reactivate --session "${CLAUDE_SESSION_ID}" → migrates state.session_id to the current HO sid (INV-002), recreates the wf-session-active.<sid> markers, opens a new session_segment in .wf-state.json (auto-heal if previous segment left open — ADR-008). Always pass $CLAUDE_SESSION_ID, never raw $WF_SID: when resuming in a new window, WF_SID may be inherited from a previous session. b. bash ${CLAUDE_PLUGIN_ROOT}/scripts/wf-orchestrate.sh <name> --query → read .sdd-state.json c. Read or.log to understand the last successful action d. Re-validate state ↔ on-disk artifacts consistency e. Re-spawn the teammates per the lookup table (via spawn_request to PM). For traceability, PM may register each teammate in the registry (bash ${CLAUDE_PLUGIN_ROOT}/scripts/wf-registry.sh add <name> <uuid> <role>) — optional since DEC-001. f. Send <resume_context> resume briefs to each re-spawned agent g. Return STATUS_REPORT to PM describing where the workflow stands

Step 6 — PM reactive loop

Identical to wf-new step 5. PM handles the workflow reactively from the resume point.

State consistency (key difference vs wf-new)

Before resuming, OR MUST verify:

• The on-disk artifacts match what the state file expects (e.g. if state = FUNCTIONAL_SPECS done, then specs.md must exist)
• No task in tasks.md is in an impossible state (e.g. DONE without TL Review = APPROVED)
• The iteration counter in review.md matches the state file

If disagreement detected → OR returns NEED_PM_DECISION with details, PM asks HO what to do (manual fix, rollback, or abort).

Rules

• Never restart from BOOTSTRAP — always resume at the current step
• Mandatory wf-check-teams preflight — fail-fast if Agent Teams flag absent
• Mandatory consistency check — prevents resuming on a corrupted state
• Silent recovery preferred — only disturb HO if the state is genuinely corrupted
• One question at a time to HO (need selection, worktree cleanup)
• Agent Teams model only — no legacy subagent model. Use TeamCreate + SendMessage exclusively.

IMPORTANT — SendMessage plain text obligatoire : le paramètre message de SendMessage n'accepte que string. Utiliser le format plain text clé: valeur — jamais d'objet {...}, jamais JSON.stringify().