Interview — constitutional-context peer conversation

What It Does

Interview reads your constitutional files — TELOS, identity, projects, system prompt, architecture — plus the CURRENT_STATE and IDEAL_STATE dimension files, checks their freshness, and confronts stale claims with observed data: Oura sleep/HRV/RHR, Conduit creation-vs-consumption time, the work registry, git cadence, and the expense ledger. The conversation opens with the sharpest claim-vs-evidence contradiction, drafts corrections you ratify, offers to populate the files the machinery can write, and closes with an ideal-state review grounded in the measured gap. On a fresh install it falls back to first-time setup. The 🎤 statusline chip (fed by InterviewDue.ts, daily launchd refresh) says when one is due; wrapping with --mark-done clears the cadence reason, and the skew/staleness reasons clear as their files are actually reviewed.

The Problem

The files that define who you are and what you're working toward drift out of date the moment you stop looking at them. A goal you set in January may be done, dead, or still right — but nobody re-reads the whole TELOS to find out, so the context the system runs on slowly rots. The usual fix is a blank "what's your mission?" prompt, which ignores everything already on file and makes you repeat yourself. This skill reads what's there first, flags only what's gone stale, and asks "still right?" instead of starting from zero.

How It Works

The skill reads the constitutional files via the freshness tooling, scores each section's staleness, and routes to the right workflow — a check-in on a populated system, or first-time setup on a fresh one. Staleness is a priority signal, not a failure flag.

Workflow Routing

Workflow	Trigger	File
ContextCheckin	default /interview on a populated system; "context check-in", "telos check-in", "what's stale", "how are we doing on…", "still on…", "review context"	Workflows/ContextCheckin.md
Phase0Setup	fresh install; DA name still reads "LifeOS"; PRINCIPAL_IDENTITY still reads "User"; PROJECTS sample-row only; .env missing required keys	Workflows/Phase0Setup.md
TelosCheckin (deprecated stub)	back-compat for explicit "telos checkin" routing	Workflows/TelosCheckin.md (redirects to ContextCheckin)

Routing decision (run before either workflow):

bun ~/.claude/LIFEOS/TOOLS/InterviewScan.ts --json | jq '[.targets[] | select(.phase == 0 and .completeness_score < 80)] | length'

• > 0 → run Phase0Setup first, then ContextCheckin.
• 0 → run ContextCheckin directly.

Quick Reference

• The TELOS is on file. Read it before asking. Generic "what's your mission?" prompts are forbidden when TELOS is populated.
• Staleness is information, not failure — a 95-day-old Goals section might still be right; the prompt is "still right?", not "you're behind."
• Per-entry on typed-ID sections (G3, M0, P2…), section-level on prose (Current State, Sparks).
• Bump on every approved edit: bun ~/.claude/LIFEOS/TOOLS/TelosFreshness.ts --bump <slug>. Without this the staleness signal degrades to noise.
• Stop signals are sacred. "Enough" / "stop" / "later" exits gracefully. State persists in the file itself.
• ID-stability rule: G3 stays G3 even when edited or dropped; new entries get the next sequential ID.

Gotchas

• The current.json day-label trap. USER/HEALTH/current.json labels itself with today's day while its last_night block may carry the newest available sleep record, days older. Quote the evidence cache's latest_sleep_record_day, never the label.
• Not every health source is live. The evidence panel (StateEvidence.ts) computes per-source liveness at run time — quote its live/dead map, never a remembered one. Claims whose only source is dead can be asked, not checked — say so instead of implying verification.
• InterviewDue.ts --mark-done at wrap is what silences the 🎤 statusline chip. Skipping it leaves the chip nagging with the interview already done — the cadence clock reads MEMORY/STATE/interview.json, not file mtimes.
• Migration must run once before TelosCheckin works. A TELOS without YAML frontmatter (no last_updated:) returns fileUpdated: null and every section reads as stale. Run bun ~/.claude/LIFEOS/TOOLS/MigrateTelosFreshness.ts once; idempotent and content-preserving (verifies sha256 of stripped content).
• The slug is normalized: "Current State" → current_state, "Wrong (Things I've been wrong about)" → wrong, "2036 — A Day in the Life…" → 2036. Always run heading text through sectionSlug() from TelosFreshness.ts.
• Pulse caches freshness for 60s. After bumping, the next /api/telos/freshness/summary call returns the cached value until invalidation. Send /reload (POST) to invalidate the cache immediately, or wait 60s.
• TelosRenderer (GenerateTelosSummary.ts) preserves the markers. It splits by ^## headings; the per-section HTML comments live inside the section body and are not re-emitted in PRINCIPAL_TELOS.md. Safe to run after edits.
• The scanner shares the freshness reader. InterviewScan.ts calls readTelosFreshness() once at startup and adds age_days, threshold_days, stale to every Phase 1+ target row. Stale sections get a +200 priority bump so they naturally rise in --next output.
• Voice notifications are "only on actual writes." Don't voice-confirm every prompt — only after a real Edit lands. The voice channel is a low-frequency signal; preserving that is what makes it land when it matters.

Examples

• "/interview" on a populated system → routing probe via InterviewScan.ts, then ContextCheckin walking the stalest sections first ("G2 is 95 days old — still right?").
• "what's stale?" → freshness summary from TelosFreshness, prioritized list, no edits unless approved.
• Fresh install (DA name still "LifeOS", PROJECTS sample-row only) → Phase0Setup first, then ContextCheckin.

Related

• /Telos — edit a single TELOS section directly (without the conversational walk).
• /Migrate — intake content from other sources (one-shot classification, not an interview).
• an identity-profile skill — manage PRINCIPAL_IDENTITY directly.
• Skill("ISA") — interview an ISA (different artifact, different workflow).