harvest

name

harvest

description

Pain-point → workflow codifier. Scans a project's session history for recurring friction patterns and emits draft workflow proposals — skill stubs, command outlines, or settings tweaks — ready to become real plugin components. Always uses an adversarial reviewer to filter one-off complaints from genuine repeating pain. Invoke when the user says 'harvest pain points', 'what should we automate', 'find workflows to codify', 'this keeps happening — make it a skill', 'this keeps coming up', 'what are we doing manually that should be automatic', 'what keeps going wrong', 'what should I turn into a skill', 'look at my session history for patterns', 'review my workflow for friction'. Accepts --project <path>; defaults to cwd. Output is advisory — no skill files are created automatically, only proposals routed through taskmaster:add-idea.

harvest — Pain → Workflow

Goal: surface things you keep doing manually and propose how to bake them into commands or skills. Different from retro in that the output is prescriptive (here is a workflow stub) not descriptive (here is what happened).

Why harvest does transcript grep in v1 (while retro doesn't): harvest's job is pattern detection across sessions, which requires raw session signals (user corrections, repeated failures) that don't surface in handovers. Retro is descriptive synthesis — handovers + lessons + git log are sufficient. The cost guard is the --window <days> cap (default 14) on transcript scan.

Parameters

Param	Default	Values
--project <path>	cwd	Same resolution as retro.
--window <days>	14	Look-back window.

SCAN Phase

1. The same artifact sources retro reads at standard depth:
   • <project>/.taskmaster/backlog.yaml
   • <project>/.taskmaster/handovers/*.md (last 7 days)
   • <project>/.taskmaster/lessons/*.md
   • <project>/.taskmaster/issues/*.md (status changed in window)
   • <project>/.taskmaster/ideas/*.md (created in window)
   • git log --since="<window> days ago" --oneline and --stat
2. Plus: targeted transcript grep on ~/.claude/projects/<project-slug>/*.jsonl for the window. Use Glob to enumerate, then Grep with these signal patterns (run in parallel):
   • "^(\\s*)(no|stop|don't|actually)\\b" — user corrections.
   • "(error|failed|fail)\\b.*\\b(retry|again|try)" — repeated failures.
   • "manually" and "every time" — explicit pain markers.
   • Tool-call sequences that repeat ≥3 times within a session (count via grep + post-processing).
3. Compress the grep hits into a digest: pattern → count → 1-line example. Cap at ≤2000 tokens.

ANALYZE Phase (proposer sub-agent, Opus)

Dispatch via Agent tool with model: "opus". Prompt template:

Before dispatching, substitute <scan output> in the template below with the token-capped digest produced in SCAN step 3.

You are the proposer for a pain-point harvest. Given the scan digest below, identify RECURRING friction patterns (≥2 occurrences across distinct sessions) and propose concrete workflow stubs to address them.

# Scan Digest
<scan output>

Return ONLY this structure:

## Recurring Pain Patterns
- pattern: <one line>
  occurrences: <count>
  sessions: <SHA / session-id refs>
  current_workaround: <how user currently handles it>

## Workflow Proposals
- title: <Verb-Noun>
  kind: skill|command|setting|hook
  pattern_addressed: <ref to one of the patterns above>
  draft: <~5 line outline of what the skill/command would do>
  blast_radius: small|medium|large
  effort: <S/M/L>

Filter out one-off complaints — patterns must repeat. Cap at 6 proposals.

ADVERSARIAL Phase (always for harvest)

Dispatch a second sub-agent via Agent tool with model: "opus". Pass ONLY the proposer's findings (not the scan digest). Prompt template:

Before dispatching, substitute <proposer output> in the template below with the proposer sub-agent's full structured response. Do NOT include the scan digest — adversarial blindness invariant.

You are an adversarial reviewer. The findings below claim to identify recurring pain patterns and propose workflow stubs. Challenge each one:

- Is this really recurring, or one bad session repeated in summary?
- Does the proposed workflow duplicate something that already exists (check by reading `plugins/*/skills/*/SKILL.md` and `plugins/*/commands/*.md` in claude-tools)?
- Is the proposal vague or actionable?
- Is the blast radius justified?

# Findings to Challenge
<proposer output>

Return ONLY this structure:

## Kept
- title: <as-is>
  reason: <one line>

## Rejected
- title: <as-is>
  reason: <one line — duplicate / one-off / vague / scope>

## New Proposals
- (same shape as proposer's Workflow Proposals — only if the adversarial sees patterns the proposer missed)

SYNTHESIZE Phase

In main context (Opus):

• Invariant: do NOT dispatch the proposer sub-agent again after seeing adversarial output. If the kept set is too small, accept the smaller set. The blindness invariant is what makes the adversarial layer meaningful.
• Merge Kept + New Proposals. Drop Rejected.
• Cap final list at 6.
• Assign confidence: high if both agents endorsed, medium if proposer-only, low if adversarial-only.

EMIT Phase

1. Report: write to <project>/.taskmaster/harvests/YYYY-MM-DD-harvest-<slug>.md if <project>/.taskmaster/ exists. Otherwise fall back to <project>/.reflect/harvests/YYYY-MM-DD-harvest-<slug>.md. Create the directory if missing. Sections: Header → Recurring Patterns (with sessions/SHAs) → Workflow Proposals (with drafts, blast radius, confidence) → Adversarial Rejections (so the user can audit the filter) → Open Questions.

2. Route proposals: every kept proposal becomes mcp__plugin_taskmaster_tm__backlog_idea_create tagged reflect:harvest. The draft outline goes in the idea body so a future task can pick it up.

3. Final message: one paragraph + report path + idea IDs.

What this Skill Does NOT Do

• Does not create skill files, command files, or settings. Output is advisory.
• Does not modify TARGETS.md.
• Does not transcript-grep beyond the configured window (cost guard).