cmb

name	cmb
description	Cosmic Microwave Background — emit a copy-pasteable handoff snapshot capturing this session's residual signal (open atomic questions, runners in flight, operator preferences) so a successor session can pick up the work without re-asking.
user_invocable	true
allowed_tools	["Bash"]
metadata	{"openclaw":{"emoji":"🌌","requires":{"bins":"[Truncated]"}}}

/cmb is a semantic handoff: it produces a human-readable Markdown artifact capturing the meaning of the current session's unwritten state — atomic questions still open, runners in flight, operator preferences active only in this session's mental model. The harness /handoff skill is infrastructural: it spins a fresh runtime continuation from the hook. The two compose. Run /cmb to draft the bridge, then /handoff (or a manual session restart) to cross it. The metaphor: a Claude Code session radiates a low-frequency residual signal — the state-not-on-disk — that is invisible unless deliberately captured. /cmb is that capture.

When to use

• Long session approaching context compaction; about to spin a successor.
• Mid-session pivot between galaxies (cosmon → mailroom → smithy).
• Cross-galaxy delib chain — child session needs the register of the parent deliberation, not just its synthesis.
• Operator suspects oral-only context will be lost (decisions tranchées, ad-hoc preferences not yet inscribed in MEMORY.md).

What the skill does — workflow

1. Run ~/.claude/skills/cmb/cmb.sh and print its stdout verbatim.
2. Do not modify the output. The header line ⚠️ REVIEW BEFORE PASTE is load-bearing — the operator's eye is the final filter against leaked reasoning being pasted into a successor.
3. Tell the operator: copy the block, paste into the successor session as the first user message. Trim sections that don't apply.

~/.claude/skills/cmb/cmb.sh                          # intra-galaxy (V0.5)
~/.claude/skills/cmb/cmb.sh --target=mailroom --intent=review   # cross-galaxy (V1')
~/.claude/skills/cmb/cmb.sh to mailroom --intent=none           # positional alias

That's the whole gesture. The script handles galaxy detection, session discovery, deterministic extractors, and the atomic-question heuristic in pure shell (no Python, no LLM call).

Cross-galaxy routing (V1')

When the snapshot is destined for another galaxy's session, two flags flip the script into cross-galaxy mode:

• --target=<galaxy> — name of the receiving galaxy. Validated dynamically against /srv/cosmon/*/.cosmon/config.toml; unknown names exit non-zero with a close-match suggestion and the full known list. to <galaxy> is a positional prose alias for the same flag.
• --intent=<verb> — required when --target is set. Closed verb set:
  • pick-up — start work on the snapshot's referenced artefacts
  • review — read and produce feedback
  • extend — add to an existing artefact (delib, ADR, doc)
  • respond — reply to an open atomic question
  • inform — context-only, no action expected
  • none — explicit honest exit (synonymous with omitting intent; valid when emitter genuinely has no verb in mind — « reste honnête »).
  • merge is explicitly forbidden. Merge-before-dispatch (CLAUDE.md / ADR-016) makes cs done the only path to fusion. A cross-galaxy merge intent would create a bypass — the receiver might fuse work that has not been validated by the source galaxy's cs done gate. The script exits non-zero with that rationale.

When --target is set, the script:

1. Adds to: <galaxy> and intent: <verb> to the frontmatter — re-internalizing the routing information that copy-paste otherwise removes (delib-20260509-5ce6 §3, D2).
2. Gates source-galaxy noise sections. Omitted from cross-galaxy output:
   • ### Working tree (git status of source)
   • ## Runners in flight (cs ensemble — source-fleet specific)
   • ### Active tmux sessions (system-wide, off-galaxy noise)
   • Galaxy-local entries in MEMORY.md (those whose file slug starts with project_ — pragmatic prefix rule; future operator-prefix index markers can refine this when the convention lands).
3. Universal sections stay — the ⚠️ REVIEW BEFORE PASTE lede, ## Atomic questions OPEN, ## Decisions tranchées (candidates), global MEMORY.md preferences, and the autopilot kill-switch state.

The frontmatter under --target extends the V0.5 4-field cap with two routing fields (to:, intent:) — not descriptive sprawl but the re-internalized path information that copy-paste removes (synthesis §3 D2). The ADR amendment lives in sibling task task-20260509-3ba1, not here.

What the skill captures

Mechanically-extracted, deterministic:

• Galaxy + worktree + branch (walk-up .cosmon/config.toml, then git).
• Recent commits (git log --oneline -5 --first-parent).
• Working tree (git status -sb).
• Live workers in fleet (cs ensemble --json, only when in a cosmon worktree; degrades to _(not in cosmon)_ outside).
• Active tmux sessions (system-wide).
• Autopilot kill-switch state (~/.cosmon/autopilot.off).
• Operator-preference index from the project-level MEMORY.md.

Heuristic, may include false positives (favored — operator prunes on paste):

• Atomic questions OPEN — assistant text messages ending with ? whose next user reply is not a short affirmative (oui, non, ok, d'accord, merci, …) are tagged open. If no user reply at all, tagged no_reply (the agent likely moved on with tool calls). Capped at 5 most-recent.
• Decisions tranchées (candidates) — V0.5. User messages classified by three orthogonal signal patterns, first match wins:
  • verb — strong, low-ambiguity decision phrases (gardons, on garde, acté, on tranche, on retient, we keep, we ship(ped), ship it, let's ship, let's keep, merge it, on (le) merge).
  • mode-feedback — enduring-rule patterns (don't, never, always, from now on, désormais, à partir de maintenant, ne fais pas, par défaut, default to, fais toujours/jamais).
  • short-reply — user message between 3–19 words, immediately after an assistant ?, that is not a pure verdict (oui/non/ok/…). Counter-questions (? at end), system-wrapped messages ([Request interrupted...], slash-commands, image-only), and long pastes (>800 chars) are filtered out. Capped at 10 most-recent.

What V0/V0.5/V1' do NOT do — by design

• No auto-write to disk. Output goes to stdout only. The agent-to-agent direct-write case (e.g. session-A writing to <Y>/docs/cmb/inbox/) uses the session-A Write tool directly; that path is a separate concern and is not part of this skill.
• No state mutation. No cs nucleate, no cs evolve, no cs done.
• No hook, no --blocking-hook. Manual gesture only — typing /cmb is the meaningful moment (ADR-091 §6 forbids blocking hooks).
• No outbox directory. ADR-091 §6 forbids per-galaxy outbox state.
• No read_seal, no acknowledgment-by-writeback. V2-if-observed.
• No merge verb in --intent. Merge-before-dispatch (CLAUDE.md / ADR-016) makes cs done the only path to fusion; a cross-galaxy merge intent would create a bypass.
• No verdict-payload pairing. When the operator gates an agent recommendation with oui (pure verdict, filtered out), V0.5 does not surface the agent's preceding proposition. That richer pairing is a future V0.6 transformation.
• No verb-form disambiguation. "On fait ça?" (question) and "on fait ça." (decision) share lexical form; V0.5 drops the verb family entirely rather than guess. Same for bare French toujours/jamais (status vs rule), kept only in verb-led forms (fais toujours/jamais).

Discovery — finding the current session JSONL

The script probes both layouts:

1. ~/.claude/projects/<encoded-cwd>/<session-id>.jsonl (Claude Code convention; the directory name encodes $CWD with / replaced by -).
2. ~/.openclaw/agents/<agentId>/sessions/<session-id>.jsonl (older openclaw convention).

For each candidate, it reads the first non-null cwd field from the JSONL and keeps the one whose cwd is the longest prefix of $PWD (with mtime as tiebreak). Same gesture as discovering an Obsidian vault by walking up.

Risks and mitigations

• Heuristic noise — false positives possible. Question heuristic capped at 5; decision heuristic capped at 10. Operator prunes on paste.
• Session-discovery brittleness — if ~/.claude/projects/ layout changes in a future Claude Code version, the script falls back to ~/.openclaw/agents/, then degrades cleanly with a warning. The deterministic extractors (git, cs) still work.
• Operator paste-and-forget — first line of the output is ⚠️ REVIEW BEFORE PASTE — auto-extracted, may include leaked reasoning. Same gesture as a release-note draft: never ship without a human eye.
• JSONL schema drift — Claude Code stores plain user text as a string on .message.content, but tool results, image messages, and assistant blocks use an array. The texts_of normalizer handles both shapes; if a future schema introduces a third form, the heuristics degrade to empty rather than crash.

Empirical validation (V0.5)

Pass run on five recent conversational sessions (cosmon, mailroom, smithy, workshop, tenant-demo), full-session scan with the V0.5 extractor:

Session	Candidates	TP	FP	Pattern that triggered
cosmon main	0	0	0	(decisions gated by pure oui — by design unhandled)
mailroom	0	0	0	—
smithy	0	0	0	— (long Element-paste correctly filtered by 800-char cap)
workshop	1	1	0	verb on shipped (relayed cmb V0 status, real meta-decision)
tenant-demo	0	0	0	—

Empirical precision = 1/1 (small N — single hit). False positives that earlier appeared in pre-tightening drafts (toujours rien, Verdict toujours vide, on fait ça? inside parenthetical questions) are eliminated by the regex tightening and the 800-char filter.

V0.5 is intentionally precision-biased: low recall is acceptable because the operator can manually fill the section, but each surfaced candidate should not require deletion. The verb-form ambiguity ("on fait ça?" vs "on fait ça.") is unsolvable in pure regex; V0.5 drops the family rather than guess.

Empirical validation (V1' — --target / --intent)

Run /cmb --target=<G> on at least 3 conversational sessions and tabulate. Pass = no regression on the V0.5 column ; new --target column shows correct gating (Runners-in-flight / Active-tmux / Working-tree absent; project_* MEMORY entries filtered; to: + intent: present in frontmatter).

Session	--target	sections gated correctly?	intent validation triggered?
cosmon worker task-20260509-b621	mailroom (--intent=review)	✅ Runners / tmux / Working-tree absent ; 0 project_* MEMORY leaks ; to: + intent: present	✅ unknown verb → exit 2 ; merge → exit 2 with merge-before-dispatch rationale ; missing intent → exit 2
(operator session #2)	to fill on next cross-galaxy use
(operator session #3)	to fill on next cross-galaxy use

Rejection-path smoke tests (all exit 2):

• cmb --target=unknown-galaxy → close-match suggestion + galaxy list.
• cmb --target=mailroom (no intent) → "intent required" + verb list.
• cmb --target=mailroom --intent=merge → merge-before-dispatch rationale + verb list.

Genealogy

• Idea: idea-20260508-819f (parent capture + feasibility + plan).
• ADR: /srv/cosmon/cosmon/docs/adr/091-cmb-handoff-pattern.md (the pattern, galaxy-agnostic).
• V0: task-20260509-1ca4 (skill skeleton + atomic-question heuristic).
• V0.5: task-20260509-dea0 (decisions-tranchées extraction).
• V1': task-20260509-b621 (--target / --intent cross-galaxy routing — this task). Parent deliberation: delib-20260509-5ce6. Sibling follow-ups: task-20260509-3ba1 (ADR amendment), task-20260509-7d27 (operator-decision doctrine), task-20260509-16f9 (doc-pédagogique).