fix(agent): reconciler change-detection + self-trigger break + failure backoff — the P0 engine (agent#187, 0.2.5-rc.2)

[unforced]

Fixes agent#187 — the reconciler was the ENGINE of the 2026-07-02 host-wide port-exhaustion outage (team vault Log/2026-07-02-port-exhaustion-incident). When auth broke underneath the daemon (a rogue loopback listener served an empty JWKS — hub#737), three compounding reconciler defects turned the 60s poll into a continuous request storm: ~15K instantiated "uni" lines + ~40K tag-schema PUTs, and two loops that 401'd 859× and 444× with zero widening.

This lands all three fixes.

1. Change-detection memo — kills the no-diff re-instantiation churn

instantiate() / instantiateThread() now compute an instantiation fingerprint of only the fields that affect instantiation (the resolved spawn spec + declared wants: + a thread's agent_status), deliberately excluding every derived/bookkeeping field the module or worker stamps back (status, pending, usage, session, turn_count, last_turn_at, …). An unchanged, already-live agent takes a fast path that skips the channel teardown+rebuild, spawn re-register, grant reconcile, and the instantiated log. It still re-resolves status via resolveStatusWithGrants so a hub grant-approval flip (pending→enabled — a change that never edits the note) keeps propagating on the poll. A real edit (system prompt / backend / mode / model / wants) changes the fingerprint and re-instantiates. Tag schemas are declared once per (origin, vault) per daemon lifetime instead of on every channel rebuild (VaultTransport guard).

2. Self-trigger loop break — diff-before-write

Each instantiate's patchStatus wrote the def note → fired the def-watch EDIT trigger on agent/definition → webhook → reload → instantiate → patchStatus → … self-sustaining. maybePatchStatus() now writes only when the resolved status/pending actually differ from what's already on the note. A status-only echo writes nothing → no updated event → the trigger never fires → the loop is dead. Verified: a status-only patch schedules no further reload (test).

3. Backoff + circuit breaker — failure loops widen instead of hammering

New shared src/backoff.ts (exponential base→cap with jitter, injectable clock/RNG) gates the three loops the incident named:

• runner loadJobs (runner.ts:tick) — the scheduler no longer re-hits a failing vault every 30s.
• def-vault listing inside loadAll (per-vault breaker) — the 401-loop the incident cited (859×/444×).
• agent-def poll (daemon.ts) — outer belt for a loadAll that throws.

Each widens on consecutive failures (60s → 2m → 4m → … → 30m + jitter), resets on success, and logs a line when the breaker opens/recovers so operators see it. Config via PARACHUTE_AGENT_BACKOFF_BASE_MS / PARACHUTE_AGENT_BACKOFF_CAP_MS, sane defaults, no new required config. (runNow is intentionally not gated — it's an explicit operator action, not the failing loop.)

Correctness notes

• The fast path preserves grant-approval propagation (status is re-resolved every pass; only the expensive rebuild is skipped).
• The fingerprint memo is cleared on every teardown path (deregisterByNote, evictOtherLiveByName) so a re-created note re-instantiates fully — audited: both this.live.delete sites clear it.
• The once-per-vault schema guard marks on success only, so a transient boot failure still retries; the tag-both write floor remains the correctness fallback.

Tests

• Change-detection both directions: unchanged→skip, content-edited→re-instantiate, wants-changed→re-instantiate, status-only note change→skip.
• Status diff-before-write: equal→no PATCH (no trigger), differ→exactly one PATCH, second pass stays quiet.
• Backoff progression on a fake clock (60s→2m→4m→…→cap, jitter bounds, ready()/succeed() semantics) + env parsing/clamping.
• Runner breaker skips ticks during the cooldown, then recovers.
• Once-per-vault schema guard: dedup across rebuilds + retry-after-failure.

Gates

• bun run typecheck — clean (exit 0)
• bun test ./src — 1475 pass / 0 fail (was 1456; +19 new)
• biome check . — 0 errors (2 pre-existing any warnings, unrelated)
• Version bump 0.2.5-rc.1 → 0.2.5-rc.2

Do not merge — reviewer pass pending (per governance). Daemon NOT restarted (code only; the live daemon is healthy).

🤖 Generated with Claude Code

https://claude.ai/code/session_01XLZtmuSs1RirWGMGyCB1QB