perf(reconcile): skip envelope on steady-state sessions, keep bootstrap safe

[JD2005L]

Summary

The SessionStart reconcile hook (hooks/reconcile-crons.sh) emits a ToolSearch + CronList + CronCreate envelope on every session start. On any already-bootstrapped workspace that envelope is pure overhead — the agent runs the ritual, finds every expected harnessTaskId already live, and prints recreated=0 adopted=0 alive=N. Several hundred ms of blocking tool calls on every REPL open, for work that meaningfully happens only on the very first session.

This PR adds a narrow fast-path: if every active registry entry already has a populated harnessTaskId, the hook exits 0 without building or emitting the envelope. If anything is unbootstrapped or a migration offer is pending, behavior is identical to today.

Why this is safe

The fast-path condition is driven entirely by existing state — no new marker files, no flag that can get out of sync:

Scenario	Registry state	Fast-path?	Result
First-ever boot	Entries just seeded, harnessTaskId null	❌ falls through	Envelope emits, agent bootstraps crons. Identical to today.
Upgrade from older version	Existing entries have populated harnessTaskId	✅ taken	Immediate fast-path. Zero regression, zero surprise.
Steady-state session	All live	✅ taken	Hook exits fast. Save the boot-time tool calls.
External cron deletion	Writeback clears stale harnessTaskId	❌ falls through	Envelope re-emits to reconcile.
Cron deleted, writeback missed it	harnessTaskId still populated but harness lacks it	✅ taken this session	Heartbeat skill's Cron reconcile step catches drift within 30 min.
Pending migration offer	MIGRATION_NEEDED=1	❌ forced through	Migration question still emitted as before.

The worst-case window for unreconciled drift is bounded at 30 min by the heartbeat skill's reconcile step (HEARTBEAT.md: "Cron reconcile — verify live CronList matches memory/crons.json..."). For context, today's behavior runs reconcile every time the user opens a new session — often hours or days apart on a long-lived agent — so this PR actually tightens the guarantee for workspaces that don't session-start often.

Implementation

Single insertion in hooks/reconcile-crons.sh, placed after migration detection (step 6) and before expected-set build (step 7). 26 lines total including comment explaining each of the 6 scenarios above. No other files touched.

if [[ $MIGRATION_NEEDED -eq 0 && -f "$REGISTRY" ]]; then
  UNBOOTSTRAPPED=$(jq '[.entries[]
      | select(.paused == false and .tombstone == null
               and (.harnessTaskId == null or .harnessTaskId == ""))]
      | length' "$REGISTRY" 2>/dev/null || echo 1)
  if [[ "$UNBOOTSTRAPPED" == "0" ]]; then
    exit 0
  fi
fi

The || echo 1 fallback is deliberate: if jq fails unexpectedly (corrupt registry, syntax bug in a future edit), we fall through to the envelope path rather than silently suppress it. Fail-safe toward doing the work, not toward skipping it.

Test plan

• Fresh workspace (no memory/crons.json pre-existing): hook emits envelope, agent creates default crons, subsequent session takes fast path
• Existing workspace with populated harnessTaskId: hook exits immediately, no envelope
• IMPORT_BACKLOG.md + unanswered migration: envelope still emitted (migration question fires)
• Corrupt crons.json: jq fails, fallback triggers envelope emission
• External CronDelete of an adopted entry: next heartbeat's reconcile catches it

Not included

This PR is just the perf fix. A tighter self-healing loop (writeback clearing harnessTaskId on CronDelete capture) is out of scope and would be a follow-up.

🤖 Generated with Claude Code