pull

name	pull
description	Post-pull orientation — absorb what changed, scan HANDOFF and RUNNING for new context, check run cards and program.md, report to user. Use when starting a session, after a git pull, or when the user says "pull", "catch me up", "what changed", or "orient".
disable-model-invocation	true
user-invocable	true

/pull — Orient, Absorb, Align

Post-pull intake workflow. The mirror of /push. Every pull is a reorientation — absorb what changed while we were away, pick up context, align on priorities.

This is a read-heavy workflow. The only output is a concise briefing to the user.

Step 1: Pull + Diff

Pull from origin and understand what came in.

git pull
git log --oneline HEAD@{1}..HEAD    # what landed
git diff --stat HEAD@{1}..HEAD      # scope of changes

If already up to date, skip to Step 2 (the scan is still valuable at session start).

Categorize the incoming changes:

• research — models/, training/, scripts/eval*, scripts/train*, experiments/, run cards
• infrastructure — scripts/ (non-training), data/, configs/
• onchain — solana/, crank/
• site — site/
• docs-only — docs/, *.md only

Step 2: Spawn Scanners

Launch two subagents in parallel to scan the living documents. These documents grow over time — the primary agent should NOT read them directly. The subagents read, filter, and return concise digests.

Scanner A: HANDOFF Scanner

Spawn an Explore agent with this prompt:

Read docs/HANDOFF.md. Find all entries that are new since [last pull timestamp from .claude/last_pull.json, or "the last 7 days" if no state file exists].

For each new entry, extract:

• Who → Who: direction of communication
• Summary: 1-2 sentence summary of what changed
• Action items for us: anything we need to do, review, or respond to
• Interface contract changes: any shared boundary changes flagged

Also check: are there any unanswered review requests or open questions directed at us?

Return a structured digest. Be concise — the primary agent will relay this to the user.

Scanner B: State Scanner

Spawn an Explore agent with this prompt:

Scan the project state for changes and current status. Check these files:

1. docs/RESEARCH-LOG.md — Read the 3 most recent entries. Summarize: what was worked on, key findings, open items.

2. Run cards (docs/run-cards/e*.md) — Read the YAML frontmatter of ALL run cards. Report:

   • Any cards with status: running (active experiments)
   • Any cards whose status changed recently
   • Any cards with rollout_coherence values filled in (completed evals)
   • Count by status: how many proposed, running, kept, discarded

3. program.md — Check git log for recent modifications: git log --oneline -5 -- program.md. If modified since last pull, summarize what changed.

4. New checkpoints or eval results — Check: ls checkpoints/ for any new directories. Check for any eval_result*.json files. Report what's new.

Return a structured situation report. Be concise.

Step 3: Update Last-Pull State

Write .claude/last_pull.json:

{
  "timestamp": "2026-03-14T12:00:00",
  "commit": "abc1234",
  "branch": "main"
}

This is used by future /pull runs to filter "what's new."

Step 4: Report

Combine the scanner results into a concise briefing for the user:

## Pull Briefing

**Incoming:** [N commits / up to date]
**Category:** [research / infrastructure / onchain / site / mixed]

### HANDOFF
[Digest from Scanner A — new entries, action items, review requests]

### State
[Digest from Scanner B — active experiments, run card status, program.md changes, new results]

### Suggested Focus
[Based on what's new: what should we work on this session?]

Keep it tight. The user wants to orient in 30 seconds, not read a report.

Step 5: Interface Contract Alert

If either scanner found changes to shared boundaries (encoding.py, wire format, SDK types, JSON frame format), call this out explicitly at the top of the briefing with a warning. These need attention before we start writing code that depends on the old contract.