Beam Agent Skill

Beam

Use when the user explicitly asks to beam, publish, or share the current local coding session with an authenticated OpenClaw receiver.

Contract

• Treat invoking this skill as approval for one snapshot upload to the named destination.
• Never upload raw JSONL, reasoning, system/developer prompts, tool output, environment values, credentials, browser state, cookies, or local paths.
• The helper is self-contained: it resolves and parses exact Claude Code or Codex transcripts, then uploads only locally redacted user/assistant messages and compact tool counts.
• Beam is a read-only projection. Do not connect a node, expose a terminal, resume the local harness remotely, or grant the receiver filesystem/tool access.
• Require HTTPS except for loopback development endpoints.
• Authenticate with BEAM_ACCESS_TOKEN, BEAM_AUTH_TOKEN, or an interactive cloudflared access login. Never print those values.
• Live synchronization is opt-in persistent behavior. Do not install hooks unless the user explicitly asks for live Beam updates.

Snapshot

Set or obtain the destination receiver URL first:

export BEAM_ENDPOINT="https://example.internal/api/v1/beam/sessions"

Prefer the harness-specific exact identifier. Resolve the directory containing this SKILL.md as BEAM_SKILL_DIR first.

Claude Code exposes the current session id:

node "$BEAM_SKILL_DIR/scripts/beam" publish \
  --endpoint "$BEAM_ENDPOINT" \
  --session-id "${CLAUDE_SESSION_ID}"

In Codex, resolve the directory containing this SKILL.md as BEAM_SKILL_DIR; normal shell-tool children expose the current thread id:

node "$BEAM_SKILL_DIR/scripts/beam" publish \
  --endpoint "$BEAM_ENDPOINT" \
  --thread-id "$CODEX_THREAD_ID"

Prefer exact harness metadata when available:

• Claude Code: ${CLAUDE_SESSION_ID} identifies the current local conversation; the helper finds its transcript.
• Codex shell tools: CODEX_THREAD_ID identifies the current thread.
• Hook adapters: pass their JSON to beam hook; transcript_path is authoritative.
• Exact fallback: Beam scans only the native Claude/Codex session roots for the supplied id and fails closed on no match or ambiguity.

Useful options:

node "$BEAM_SKILL_DIR/scripts/beam" publish --endpoint "$BEAM_ENDPOINT" --session /path/to/session.jsonl
node "$BEAM_SKILL_DIR/scripts/beam" publish --endpoint "$BEAM_ENDPOINT" --thread-id "$CODEX_THREAD_ID"
node "$BEAM_SKILL_DIR/scripts/beam" publish --endpoint "$BEAM_ENDPOINT" --complete
node "$BEAM_SKILL_DIR/scripts/beam" publish --endpoint "$BEAM_ENDPOINT" --dry-run

--dry-run prints the sanitized payload locally and performs no network request.

On success, return only the Beam URL and a short disclosure summary: source harness, shared message count, whether older entries were truncated, and whether the beam is complete.

Live Hooks

The helper accepts the snake_case JSON emitted by both Claude Code and Codex lifecycle hooks:

node "$BEAM_SKILL_DIR/scripts/beam" hook --endpoint "$BEAM_ENDPOINT" --quiet

Configure root Stop hooks for turn-by-turn updates. Claude Code can also use SessionEnd for finalization; Codex finalization uses an explicit publish --complete. See:

• references/claude-code-hooks.json
• references/codex-hooks.toml

Hook failures must not block the coding session. Keep them quiet on success; log one concise redacted error on failure.

Full installation, authentication, data-boundary, and hook documentation:

README.md.

Development

node --check skills/beam/scripts/beam
node --check skills/beam/scripts/beam-session.js
node --test skills/beam/scripts/beam.test.mjs
scripts/validate-skills