AI Start

Start

Bootstraps a coding session: a deterministic Python script (session_bootstrap.py) renders the whole welcome dashboard, so the agent runs one command, prints the markdown verbatim, and stops. Re-probing git/sqlite/manifests/board APIs agent-side blows the latency budget (operator-pain #18b), so the script collects every field and caches the board call — cold path <3 s with board, warm path <500 ms.

Workflow

Principles: §10.1 KISS (one deterministic command; zero agent-side re-derivation).

1. Run exactly this argv — literal, no flags moved, no shell added:

uv run python .ai-engineering/scripts/session_bootstrap.py --format=markdown

2. Print its stdout verbatim and stop. That is the whole skill.
3. This exact argv is enrolled in the trusted-script lane (hooks-manifest.json trustedArgvs, D-131-12) so it bypasses RTK rewriting + IOC re-evaluation. Any other form (reordered flags, plain python3, missing --format) falls back to the full IOC path and degrades latency.

Hard rules

• Do NOT read the manifest, run git, query sqlite, hit gh, glob the skills/agents tree, or count LESSONS.md from the agent side — the script already embedded all of that in the markdown.
• Do NOT rewrite the emitted markdown. The format is the cross-IDE contract (Claude Code, Codex, Antigravity, Copilot render the same bytes).
• Do NOT invoke /ai-session-watch from here. Observation is always-on via PreToolUse+PostToolUse hooks (instinct-observe.py), consolidated at session end by the Stop hook (instinct-extract.py). The dashboard surfaces an N to review CTA when the unconsolidated backlog exceeds observations/meta.json deltaThreshold; operators then run /ai-session-watch --review manually.

What the dashboard already contains

Trust the emitted markdown — do not re-render any of these agent-side:

• Project identity: CONSTITUTION mission as tagline.
• Stack posture: surfaces.enabled · gates.mode (one line, so layer drift is obvious).
• Counts: skills, agents, lessons, active decisions, accepted risks, recent_events_7d.
• Active work: spec id + state + title, plan status (incl. shipped-pending-pr-merge per plan-schema.md), task progress.
• Recent commits: last 5 SHA + subject.
• Recent lessons: last 3 ### headers from LESSONS.md with a gist line (context prefix stripped server-side).
• Board: full per-status breakdown via paginated GraphQL (no sample-size truncation).
• Compatibility: ### ⚠ Compatibility block appears only when the manifest deviates from defaults (today: gates.mode != regulated).

Board behaviour

• The script handles gh project item-list with a hard 4 s subprocess timeout and a stale-while-revalidate cache at .ai-engineering/runtime/board-cache.json (fresh ≤60 s, stale-allowed up to 5 min).
• On board failure the JSON sets board_summary.unavailable: true and the markdown shows board unavailable (reason) — never blocks the rest of the dashboard.

When the script is unavailable

If the script exits non-zero or the venv has no uv, fall back to a one-line banner: ai-start unavailable — repo not bootstrapped, run \ai-eng install`.` Do NOT reconstruct the dashboard by hand.

Examples

/ai-start

Runs the script, prints the dashboard (active spec, last 5 commits, board items by status, project counts, quick-action chips), stops. Mid-session re-bootstrap after /clear uses the same single command; a fresh board cache makes it near-instant.

Integration

• Called by: user directly; IDE instruction files (FIRST ACTION mandate per CONSTITUTION).
• Calls: session_bootstrap.py --format=markdown (only).
• Does not call: /ai-session-watch, /ai-board discover, manifest readers, or any other skill. Suggestions (e.g. "no active spec — run /ai-brainstorm") are embedded inside the emitted markdown.
• See also: /ai-onboard (human onboarding, different audience), /ai-branch-cleanup (pre-start hygiene).