/pulse

Read eight offline session signals and record the observation through the canonical seed writer. Use when the user requests /pulse, a quick session health check, or a boundary readout. The underlying stdlib script calls no model, network service or test suite.

Arguments

• $ARGUMENTS — optional --boundary, --topic, --source or --signals-file options for session-pulse.py. Pass --topic only for a fleet topic, consistently to reader and writer.
• --source names the actual host: claude, codex or chroxy. Leave it unspecified when unknown. Never consult Claude's meter for Codex/chroxy; the Claude arm additionally requires the real $CLAUDE_CODE_SESSION_ID.

Instructions

1. Read, do not manufacture measurements. The script reads the canonical seed and local git state. Tests and CI are cached STATE values, never fresh runs. Unknown measurements are n/a, including open counts and turns when no explicit cached count exists.

2. Observe into a scratch snapshot. Run the installed script, adding only appropriate arguments above. Show its eight lines verbatim:

   PULSE_SNAPSHOT=$(mktemp "${TMPDIR:-/tmp}/session-pulse.XXXXXX")
   python3 ~/.claude/scripts/session-pulse.py --snapshot-file "$PULSE_SNAPSHOT"
   

   Plain invocation is observational. --snapshot-file may write a scratch JSON file; it refuses seed names, aliases and the handoff directory. It does not write a seed. The reader always exits zero, including source, timeout and snapshot-output failures, so inspect the eight lines rather than treating the exit code as proof of persistence.

3. Persist only with the real session identity. When the host supplies the actual session ID, record the snapshot using the writer:

   python3 ~/.claude/scripts/session-seed.py write \
     --refresh-pulse "$PULSE_SNAPSHOT" --session "$ACTUAL_SESSION_ID"
   

   ACTUAL_SESSION_ID is supplied by the host, never guessed from a worktree, a seed or the newest transcript. A real Claude host may omit --session and use its harness-provided environment ID. If the ID is unavailable, keep this pulse observational and report that it was not saved. Pass the reader's same --topic when applicable. Do not hand-edit the seed, even for a snapshot. The writer preserves the existing date/body/decisions, archives an incumbent owned by a different session and runs fresh canonical proof. Missing or hollow legacy seeds remain unchanged when refresh refuses. Report REFUSE or snapshot-save failure honestly; never claim the baseline was stored without seed:.

4. Act on the readout. A WARN never stops authorized work. Invoke /decide for an actual pending choice, respecting owner-reserved authority. Run /visual-brief when a pulse delta crosses a WARN threshold or at package close. End the surrounding response with the fleet four-slot **Status:** block and one-task **Next:** sentence.

Cached signals contract

--signals-file accepts only an explicitly supplied JSON object with source matching --source and a full UTC observed_at. Optional open_items, turns and compactions are nonnegative integers, never booleans. context_pct is a finite number from 0 to 100. estimate and actual are finite nonnegative numbers supplied together with their common nonempty unit. Keep their source and timestamp in the cached input; missing data stays unknown. Do not estimate usage or context from spend, elapsed time or another host.

The eight lines cover scope; seed age and dispatch; HEAD/branch/tree/ahead/behind; commits since seed/open items/turns; cached tests and CI; decisions count/age/default; usage/context/ compactions; and delta/WARN. Claude usage may come only from usage-pace.py --oneline. All subprocesses share a 0.8-second deadline; unavailable sources degrade to n/a.

The readable Claude usage line remains the helper's bounded raw --oneline output. Material comparison uses only supported printed measurements at that formatter's precision: direct sd, fh and fable percentages; total or explicitly labelled subtotal spend; fable spend; and derived capacity percentage/cap. Sample timestamps and ages, reset countdowns, rates, burn, headroom and projected landings are excluded. Missing or unrecognized measurements remain unknown. This small presentation signature reuses the single bounded --oneline call; it neither recalculates accounting nor makes another source call for a separate JSON report.

A missing previous snapshot says no baseline; unchanged material state says no material change since TIME. Changed material fields are named. Decision recurrence counts advance only when snapshots are persisted. Raw ages and recurrence counters do not manufacture material changes, but stable threshold booleans make a newly crossed WARN visible in delta and remain unchanged while the threshold remains crossed. WARN thresholds are a repeated decision text across at least two recorded pulses, more than three open items, a seed older than 24 hours, at least two compactions, a dirty working tree at a boundary, or stated estimate drift above 25 percent. Legacy pending DECISION prose warns with known age and unknown count until structured metadata is supplied.

Seed metadata contract

The full writer accepts --decisions-file containing a JSON list of {"text":"Approve the proposed plan","opened_at":"2026-10-28T13:00:00Z","default":"Hold the proposed plan pending an answer"} objects; default omission is an empty list. Carry every unresolved choice and its original opened_at. This default is a recommended plan, never reserved execution authority. --pulse-file carries a snapshot in a full seed write. --refresh-pulse updates snapshot and session metadata only, preserving date and body. Both flow through the same validation, archive and proof path. decisions and last_pulse are single-line JSON flow values within YAML frontmatter; conventional block-style YAML is outside this stdlib contract.