recall-workflow-agent

recall-workflow-agent

Overview

Every Workflow subagent's full transcript is written to disk as JSONL. To get one agent's prompt + output you locate its agent-<id>.jsonl and read it — you do NOT re-run the workflow or ask the user. The reliable locator is a distinctive PHRASE, because the readable agent label is not stored on disk.

Where it lives

~/.claude/projects/<project>/<session>/subagents/workflows/<wf_id>/
    journal.jsonl          # per agent: {type, key:<HASH>, agentId, result:<the OUTPUT>}
    agent-<agentId>.jsonl  # per agent: full transcript (prompt = its user msg; + turns)
~/.claude/projects/<OTHER-project>/<session>/workflows/scripts/<name>-<wf_id>.js   # the script (holds the labels)

Gotcha: the transcript dir is under the SESSION's project, which is frequently a different <project> than where the .js script lives. Never assume co-location — a phrase grep finds it wherever it is.

Do this — the bundled tool locates + extracts + saves the files

S=~/.agents/skills/recall-workflow-agent/scripts/extract_workflow_agent.py
# best: a phrase you SAW in the output, or a phrase from the prompt/lens task:
python "$S" --phrase "Vaporware-filtered maturity table" --slug decision --print
# don't know a phrase? list every agent in the run (previewed by its report HEADLINE), then pick:
python "$S" --workflow multiai-orchestration-research --list
# scope a phrase to one run if it's ambiguous:
python "$S" --workflow <name> --phrase "<distinctive>"

Prompt is saved to <slug>-PROMPT.md, output to <slug>-OUTPUT.md. The tool takes the output from the journal result (the authoritative captured return), so it's clean even for tool-using agents.

Identify by (most→least reliable)

1. A phrase from the OUTPUT you saw in the UI — unique, best.
2. --list the workflow's agents (each previewed by its report headline) → pick.
3. A phrase from the PROMPT / lens task — works, but read the synthesizer gotcha below.

Common mistakes

• Grepping the agent label (synthesize:multiai-decision): it is NOT in the transcript or journal — journal stores a hashed key; labels live only in the .js script. Grep a content phrase instead.
• Searching only under the script's project dir → 0 files. Grep across ALL of ~/.claude/projects (the tool does).
• Fanning out the Grep TOOL across ~/.claude/projects: on Windows it can die with ENAMETOOLONG (giant argv over thousands of files). Use the bundled tool (Python glob + line scan) or rg via Bash — not the Grep tool.
• A lens phrase matching two agents: a SYNTHESIZER's prompt embeds every upstream agent's report verbatim, so a lens name/tag hits the lens agent AND the synthesizer. Use a phrase from that lens's OWN narration/output, or --list and pick by preview. The tool reports every match so you can narrow.
• Decoy — your OWN current session's subagents: Agent-tool subagents live at subagents/agent-*.jsonl (NO workflows/<wf_id>/ layer). If your task text quotes the workflow name/phrase they self-match. The tool's glob requires .../subagents/workflows/<wf_id>/agent-*.jsonl, so it excludes them — a manual grep won't.
• Treating joined assistant turns as "the output": tool-using agents interleave narration between tool batches. The real output is journal result (final return) — what the tool uses.
• content is a string OR a list of {type:text} blocks — handle both (the tool does).