interpulse

What it does: Single-hook session context-pressure monitor. A PostToolUse hook (matcher Edit|Write|Bash|Task|NotebookEdit|MultiEdit, 5s timeout) maintains ephemeral state in /tmp/interpulse-${SID}.json and computes a dual-threshold level = max(pressure_level, context_level). Pressure is heuristic: +1.0 per call (+1.5 for Read/Grep/Task/WebFetch/WebSearch), -0.5 decay per 10 idle minutes, plus a cumulative token estimate (tool_output length / 4). Context is ground truth: it reads context_window.remaining_percentage from hook stdin and normalizes for Claude Code’s 16.5% autocompact buffer, falling back to pressure-only when absent (subagents/old CC). Warnings (yellow/orange/red) are emitted via additionalContext with 5-call debounce + severity-escalation bypass; red writes a /tmp checkpoint markdown; orange emits an interband checkpoint_needed signal intended for intermem; every call writes an interband pressure payload for a statusline. One skill (pressure) renders a dashboard from the state file. lib/interpulse-lib.sh holds the shared state + normalization helpers.</purpose_confirmed> Low conceptual overlap but it is the canonical owner of the context-pressure concern after being extracted from intercheck — intercheck README/code still references context pressure and interline’s statusline still reads the OLD intercheck/pressure channel, so ownership is split mid-migration. interstat (token-efficiency benchmarking/session analytics) and tool-time (tool-usage analytics via hooks) overlap on the raw signal (per-tool counts, token estimates) but serve post-hoc analysis vs interpulse’s live in-session gauge. intermem is the intended downstream of the orange checkpoint signal but does not consume it. interline is the intended statusline consumer but reads the wrong channel. No true duplicate; the redundancy is a half-finished extraction, not two plugins doing the same job.

Rationale: The core engine is sound and notably modern for its size: it prefers Claude Code’s ground-truth context_window.remaining_percentage and only falls back to heuristics, which sidesteps the usual naive-token-counting critique; debounce + severity-escalation-bypass is well-judged; hooks fail open (trap ‘exit 0’). This is not a deprecate — it owns a real concern (context-rot early warning, repeatedly validated in the 2026 token-economics/context-rot findings as a first-order cost/quality driver). But it has two concrete integration breaks that make its outputs invisible, plus internal doc/skill drift, and it stops at ‘warn’ when the SOTA move is to act. Hence improve, not keep: fix the broken wiring and close the pressure->action loop, but don’t rebuild the engine. Specific evidence: (1) interpulse writes the statusline payload to interband channel interpulse/pressure/, but interline/scripts/statusline.sh Layer 4 still reads intercheck/pressure/${session_id}.json — the pressure indicator never reflects interpulse. (2) interpulse emits interband interpulse/checkpoint/ checkpoint_needed at orange “for intermem,” but intermem has zero consumers of that channel — a dead signal. (3) Doc drift: README.md and AGENTS.md document a skills/status/SKILL.md and /interpulse:status, but the actual skill is skills/pressure/SKILL.md and the command was renamed to /interpulse:pressure (git commit c218f7b) to resolve a 4-way status collision; README’s threshold table (Yellow 60/Orange 90/Red 120) also disagrees with its own prose (60/75/85) and with CLAUDE.md. (4) Token estimate output_len/4 double-counts vs the now-authoritative remaining_percentage and over-counts subagent/Task output that never enters the parent context.

SOTA gap: Passive gauge with no relief action. Frontier (Anthropic write/select/compress/isolate, +29%/+39%; demand-paging for context arXiv:2603.09023; SWE-Pruner arXiv:2601.16746; structured note-taking/externalization) treats a pressure threshold as the trigger for automated compaction/eviction/externalize-to-memory. interpulse has the authoritative input and even fires a checkpoint event, but the loop is dead. Missing frontier idea: pressure-as-trigger, not pressure-as-warning.

Redundancy: It is the canonical context-pressure owner post-extraction from intercheck, but ownership is split mid-migration: intercheck still references the concern and interline still reads the legacy intercheck/pressure channel. Signal-level overlap with interstat (session token analytics) and tool-time (hook-based tool-usage analytics), but those are post-hoc whereas interpulse is the live gauge. intermem/interline are intended collaborators, not duplicates. No merge target — fix the seams instead.

Actions:

• Fix the statusline integration: either repoint interline/scripts/statusline.sh Layer 4 to read interband channel interpulse/pressure/ (with intercheck/ as fallback during migration), or coordinate the channel name — currently the live pressure indicator never reflects interpulse output.
• Close the dead checkpoint loop: make intermem (or a documented consumer) read the interband interpulse/checkpoint/ checkpoint_needed signal at orange and auto-synthesize/externalize session memory — turning the existing signal into the SOTA pressure-triggers-compaction/externalization action (cite Anthropic context-engineering write/select/compress/isolate, +39% with memory tool).
• Reconcile docs to code: update README.md + AGENTS.md to reference skills/pressure/SKILL.md and /interpulse:pressure (not status), and fix the contradictory threshold tables (README prose 60/75/85 vs table 60/90/120 vs CLAUDE.md) so the published numbers match the hook.
• De-emphasize the heuristic token estimate when context_window.remaining_percentage is present (avoid double-counting and over-counting subagent/Task output that never enters parent context); make ground truth authoritative and the heuristic a labeled fallback.