session-start

Session Start (orientation)

Reproduce the effect of the user saying "understand this project" at the start of a session: read the project's Markdown corpus, build a genuine understanding, then report a concise project summary and an honest current-status snapshot.

This is a read-and-report workflow with no side effects: do NOT export a figure deck, edit memories, write report files, or run analyses. (For a status snapshot + memory reconciliation use /sync; for figure↔outline gap TODOs use /todofig.)

Default to English for the report; keep file names and identifiers verbatim. Override the report language via the project CLAUDE.md.

Configuration

Resolve config in priority order: env var → project CLAUDE.md ## Research stack block → default.

Setting	## Research stack key	Env var	Default
Outline file	Outline file	OUTLINE_FILE	outline.md
Wiki dir	Wiki dir	WIKI_DIR	docs/wiki/ (if present)

Modes — light (default for auto-trigger) vs full

• Light mode — the default when invoked by a SessionStart hook (its injected context will say "light mode"), or when the user types /session-start light. Do Step 1 → Step 3 → Step 4 only; skip Step 2 (the full corpus sweep). Rationale: the harness already auto-injects CLAUDE.md, the project MEMORY.md, and agent memories at startup; light mode only adds the outline + wiki landing page synthesis that isn't auto-loaded. Keep the report tight.
• Full mode — when the user types /session-start (no arg), /session-start full, or passes a focus area. Do all steps, including the Step 2 corpus sweep.

If unsure which mode, default to light.

Step 1 — Read the orientation core (always — both modes)

Read these first — they are the spine of the project:

1. CLAUDE.md — the single most authoritative file (## Project context = scientific identity; ## Research stack = config).
2. The Outline file (from ## Research stack) — the canonical outline / scientific ground truth; supersedes any digest table in CLAUDE.md if they disagree.
3. Any project TODO / execution-state file the outline or CLAUDE.md points to (what is ✅ / 🟡 / ⬜ / 🚧).
4. The wiki landing page if a Wiki dir is configured (e.g. <Wiki dir>/home.md or Home.md) — one cheap file that mirrors the outline in curated form and tells you which wiki pages to open in Step 2. Scientific-content precedence: Outline file (canonical) > wiki (curated) > memory. If the wiki disagrees with the outline, trust the outline and flag it.
5. The project MEMORY.md files (usually already injected at session start — re-scan rather than re-read if so).

Step 2 — Sweep the rest of the Markdown corpus (FULL mode only — skip in light mode)

Discover every other .md in the workspace and read those relevant to current state:

find . -name '*.md' -not -path '*/node_modules/*' -not -path '*/.git/*' 2>/dev/null

Prioritize and read as needed: pipeline/method docs, the configured Wiki dir result & method pages (from the landing page's links — open the pages relevant to current work rather than the whole vault), and the most recent dated reports (e.g. under a sync_reports/ or todofig_reports/ dir if present). Budget note: read the Step 1 core fully; for the rest, read the most recent / most relevant files and skim archives. State in the report if you deliberately skipped archives.

Freshness check: if the Outline file is newer than the TODO / latest report, flag that the status files may lag the outline.

Step 3 — Build the understanding

Synthesize — do not just concatenate file contents. Make sure you can answer:

• The claim — the one-line thesis and the canonical key-finding statement.
• The narrative spine — Gap → Question → Approach → Finding → Implication.
• The load-bearing fact(s) — the empirical result(s) the argument rests on, and any invariants / non-circularity commitments.
• Result structure — the current results/figures layout and what each asserts.
• Pre-registered commitments — any success criteria + fallbacks that are not to be relitigated.

Step 4 — Report

Deliver a concise orientation report (this is orientation, not a manuscript):

# Project orientation — YYYY-MM-DD

## Identity
- Title / lead author · PI / target venue / central hypothesis (one sentence)

## Narrative spine
- Gap → Question → Approach → Finding → Implication (one line each)
- Invariants / non-circularity commitments
- Most load-bearing empirical fact

## Result structure (current N results / M figures)
- R1 … RN, one line each + owning figure + what it asserts

## Current status
| Result | Figure | State | Remaining work |
(✅ done / 🟡 partial / ⬜ not started / 🚧 new·in-progress)

## Most urgent work (P0)
- 1–3 items from the TODO's P0 tier

## Caution / drift flags
- File-to-file inconsistencies, status files older than the outline, unresolved CRITICAL issues

End with one line offering next steps.

Argument handling

• light → light mode (Step 1 → 3 → 4; skip Step 2). Also what a SessionStart hook requests.
• full or empty → full mode (all steps, full corpus sweep).
• A focus area (e.g. R3, figures) → full mode but narrow Step 4's status report to that focus — give that result/figure its full detail and compress the rest into a one-line context header.

Guardrails

• Read-only. No exports, no memory edits, no report files, no analysis runs. If the user wants any of those, point them to /sync or /todofig.
• Outline wins. When a CLAUDE.md digest table and the Outline file disagree, trust the outline and flag the discrepancy. Precedence: Outline file > wiki > memory.
• No fabrication. If a file is missing or unreadable, say so; do not invent panel contents, numbers, or status.
• Observed state over stale memory. If a memory/report contradicts the current outline or TODO, prefer the current files and flag the drift.