reflection

Triggered on-demand when the founder asks you to "reflect" inside a thread turn. It produces a structured opening report and, when the founder is satisfied, persists learnings and optional KB entries directly — no endpoints, no transcript store, no task dispatch.

Identity

Read your agent name from your workspace directory name (or the # Agent: header of your bootstrap CLAUDE.md/AGENTS.md) if you don't already know it. Every happyranch learning ... and happyranch kb ... call takes --agent <your_name> where required.

START — opening report as a thread reply

1. Find the window for the report. Scan task_history.md for your most recent reflection (look for "reflection" — or the legacy "review" — in the briefs/outcomes). If no prior reflection exists, use "all-time, capped at 30 days."

2. Gather inputs:

   • task_history.md (in the workspace root).
   • Per-agent learnings — learnings/_index.md (drill into individual LRN-NNN-<slug>.md entries dated within the window).
   • happyranch audit --org {ORG_SLUG} --agent <your_name> --since <window_start>.

3. Emit the opening report as a thread reply. Use exactly these section headings, in this order:

   Since last reflection

   Window dates and one-line counts by terminal status.

   Notable tasks

   3–5 items you picked for significance (not just most recent). One-line takeaway each.

   New learnings

   Learnings added in the window. Bullet list.

   Open questions / frictions

   Your own reflection — anything confusing, contradictory, a recurring issue pattern, a decision you're unsure about. Can be empty.

   Suggested topics

   2–3 things you think are worth discussing.

4. Wait for the founder. The rest is a normal conversation.

END — direct learning/KB capture (no endpoint)

When the founder indicates the reflection is done:

1. Summarize. In your own head, list: what was discussed, what decisions got reached, what remains open.

2. Extract learnings. A learning is a durable, non-obvious operational fact you'll want on future tasks. Skip things already recorded. For each learning, write a YAML payload:

   slug: lrn-<slug>
   title: <title>
   topic: <topic>
   tags: [tag1, tag2]
   body: |
     <markdown body>

   Then call:

   happyranch learning add --org {ORG_SLUG} --agent <your_name> --from-file /tmp/lrn-<slug>.yaml

3. Identify KB-worthy material. Prefer durable, cross-agent knowledge with an expected useful lifespan of at least 12 months; keep agent-private lessons in memory. Write /tmp/kb-<slug>.md with YAML frontmatter and a markdown body, then call:

   happyranch kb add --org {ORG_SLUG} --agent <your_name> --from-file /tmp/kb-<slug>.md

4. Post a final thread reply summarizing outcomes: what was discussed, decisions reached, learnings/KB entries written. This replaces the lost transcript.

Why single-line call + temp file

The --from-file pattern matches happyranch report-completion and other callbacks. Claude's headless-mode permission matcher treats multi-line bash as multiple separate commands, which breaks the allowlist. Staging the payload in a temp file and invoking happyranch once keeps the callback inside the Bash(happyranch *) allow rule.

No dispatch

A reflection session does not create tasks. Do not call happyranch dispatch or happyranch run. If actionable work surfaces, recommend the founder open a thread or dispatch it separately, or self-dispatch a root from a normal thread turn — not from reflection.

This is skill-level enforcement. The reflection skill has no task-path auth; there is no talk-mode dispatch surface to abuse. If you need to act on a finding, tell the founder and let them decide the next step.

What NOT to do

• Don't start tasks (happyranch run ...) from inside a reflection. If something actionable comes up, tell the founder explicitly.
• Don't write learnings you've already written — the daemon appends verbatim, so duplicates will clutter.
• Don't treat KB entries as a catch-all for in-reflection notes. KB is for durable, cross-agent-relevant knowledge. Everything else is a per-agent learning.
• Don't call happyranch talk start|end|resume|abandon — those endpoints no longer exist. This skill replaces the talk flow entirely.