rdfitted/start

/start

Session-start ritual. Establishes a learning-capture heartbeat for the duration of this REPL session.

When this runs

• Automatic: at session start, after the wiki + project-dna read described in ~/CLAUDE.md → "Session Start".
• Manual: user invokes /start.

If the heartbeat is already scheduled in this session (check CronList first), do NOT re-create it — just print the existing job id and exit.

Workflow

Step 1: Check for existing heartbeat

Call CronList. If any job's prompt contains the sentinel <<start-heartbeat>>, the heartbeat is already running for this session. Print the existing job id and stop — do not schedule a duplicate.

Step 2: Detect tier

$cwd  = (Get-Location).Path.TrimEnd('\','/')
$home = $env:USERPROFILE.TrimEnd('\','/')
$hasJsonl = Test-Path ".ai-docs/learnings.jsonl"
$hasAiDocs = Test-Path ".ai-docs"

if ($cwd -ieq $home) {
    $tier = 'tier2'    # institutional mode — target global wiki
} elseif ($hasAiDocs) {
    $tier = 'tier1'    # project mode — target .ai-docs/learnings.jsonl
} else {
    $tier = 'tier1-uninit'  # code project but no .ai-docs/ yet
}

Decision rules:

• cwd == $HOME → Tier 2 (institutional / global wiki)
• .ai-docs/ exists → Tier 1 (project)
• Otherwise → Tier 1 uninit — suggest /init-project-dna to user; still schedule the heartbeat (it will buffer to wiki/log.md until init runs)

Step 3: Load minimal context (do not re-read what CLAUDE.md already loaded)

• Tier 1: peek at .ai-docs/project-dna.md "Patterns That Work" header to confirm it's readable. Do NOT dump contents.
• Tier 2: grep ~/.ai-docs/wiki/log.md for the last 3 | curate | or | add | entries to anchor recency. Do NOT dump contents.

This step is just to confirm the targets are writable. Keep it tight — under 5 lines of output.

Step 4: Schedule the heartbeat via CronCreate

Use CronCreate with these parameters:

• cron: 17,47 * * * * (every 30 min at :17 and :47 — off-mark per fleet jitter guidance)
• recurring: true
• durable: false (session-only; do not persist to disk)
• prompt: the heartbeat prompt below (must contain the sentinel <<start-heartbeat>> so Step 1 can detect duplicates)

Heartbeat prompt (self-contained — future ticks have no memory of this session):

<<start-heartbeat>>
Heartbeat tick. Scan what's happened in this REPL session since the last tick (or session start if first tick). Identify NEW patterns worth capturing — generalizable insights, surprising failures, confirmed approaches, anti-patterns, cross-client connections.

Quiet is fine. If nothing meaningful happened, just print "heartbeat: no new patterns" and stop. Do NOT invent learnings to justify the tick.

If there ARE new patterns, capture them based on mode:

**Tier 1 (cwd has .ai-docs/)**: Append one JSONL entry per pattern to {cwd}/.ai-docs/learnings.jsonl using this schema:
{"ts": "<iso>", "session": "<short-id>", "task": "<one-line>", "outcome": "success|failed|partial", "keywords": ["tag1","tag2"], "insight": "<what was learned>", "files_touched": ["path"]}

**Tier 2 (cwd == $HOME)**: Append one line per pattern to ~/.ai-docs/wiki/log.md under today's date:
## YYYY-MM-DD | candidate flagged | <one-line pattern> (source: this session, <brief context>)

**Tier 1 uninit (code project, no .ai-docs/)**: Append to ~/.ai-docs/wiki/log.md as a buffered candidate:
## YYYY-MM-DD | candidate flagged (uninit project: <cwd basename>) | <one-line pattern>

After capture, print a 1-2 sentence summary: "heartbeat: captured N patterns to <target>". Then stop — do not continue other work.

Step 5: Print the session-start banner

Output a tight banner (4-6 lines max):

/start — heartbeat scheduled
  tier: <tier1|tier2|tier1-uninit>
  cwd: <basename of cwd>
  target: <.ai-docs/learnings.jsonl | ~/.ai-docs/wiki/log.md>
  cadence: every 30 min (:17, :47) — fires when REPL is idle
  job id: <id from CronCreate>
  expires: 7 days (CronCreate ceiling) or session end

If Tier 1 uninit, append one extra line: note: no .ai-docs/ — run /init-project-dna to enable project-level capture (heartbeat buffers to wiki/log.md until then)

Step 6: Stop

Do NOT proceed with other work. The user's actual prompt (if any) follows; respond to that next.

Rules

• Idempotent: if a heartbeat with the <<start-heartbeat>> sentinel already exists in this session, do not create a second one.
• Session-only: never use durable: true. The heartbeat dies with the REPL on purpose — fresh detection each session.
• Off-mark cadence: always use minutes like :17/:47, never :00/:30. (See CronCreate fleet-jitter guidance.)
• Quiet ticks are valid: heartbeat ticks that find nothing should print one line and stop. Do not fabricate patterns.
• Don't dump context: Step 3 confirms targets exist; it does not re-read the wiki or project-dna (CLAUDE.md already triggered those reads).
• Tier detection wins: if cwd == $HOME, ALWAYS Tier 2 even if a stray .ai-docs/ appears in the home dir — institutional mode is the contract for home sessions.

Cross-references

• /curate-learnings — Synthesizes captured heartbeat entries (Tier 1: jsonl → project-dna.md; Tier 2: wiki/log.md candidates → global wiki pages).
• /init-project-dna — Initializes .ai-docs/ in a code project (run this if /start reports tier1-uninit).
• ~/.ai-docs/wiki/log.md — Tier 2 capture target + provenance trail.
• ~/CLAUDE.md → "Session Start" — Invokes /start after wiki/project-dna reads.