v1.6.2 — notification reliability hardening

Three reliability fixes for the notification daemon, each independently adversarially gated (audit/grok + audit/gatekeeper) before merge.

• Per-session pointer-inject backoff (#32): the daemon no longer re-injects the same [NOTIFY] pointer every 3s into a session that hasn't drained its inbox. A 90s per-session backoff keyed on inbox signature stops the hammer; genuinely new messages (changed signature) still inject promptly. Fixes the keystroke-storm failure mode.
• Suffix-peer self-parent liveness routing (#33): a suffix peer (-codex/-gemini/-grok) whose explicit parent override points at itself is now ignored in favour of the suffix-strip, so stop signals route to the real supervisor. Top-level sessions still self-suppress correctly.
• Inactive-peer backstop fresh-activity check (#34): the dispatcher-page backstop now measures inactivity from the newest of last_tool_activity / last_activity / current_task.started_at, so a freshly-restarted or freshly-dispatched active peer is no longer falsely reported inactive. Genuinely stalled peers still surface.

Production-validated on the live daemon: zero false backstop firings post-deploy; 58 tests + subtests pass.

v1.6.1 — atomic_deploy self-contained clones (#5)

Patch: atomic_deploy.sh now uses self-contained git clone per version dir instead of git worktree (#5 — fixes gitdir-pointer fragility that broke redeploys after symlink-flip). No runtime/hook behavior change. Closes the last open fleet-notify issue.

v1.6.0 — Reliable worker→supervisor stop signaling

v1.6.0 — Reliable worker→supervisor stop signaling

Hardens the core guarantee: a supervisor always learns when its worker stops — whether the worker reports back, forgets, or errors out — with no human polling.

Highlights

• Always notify on worker stop (#30): removed the suppression that silently dropped a stop when the worker had no bound task and no recorded outcome. Every distinct worker stop now wakes the supervisor (with task/outcome inline when present, an honest "stopped — no current task recorded" when not). Same-stop-event idempotency prevents double-fire; a genuinely-distinct later stop fires fresh. A worker in a no-op loop now surfaces (a bug to fix), rather than being silently muted.
• Notify dispatcher on all peer stops (#27) + stale-task backstop fix (#29): the periodic backstop re-checks that a worker's current_task is still live before paging, so stale/orphaned task bindings don't produce false "still working" notifications.
• Dispatch activation validation (#28).
• Install hardening (#26): the installer AST-scans all imports (top-level + lazy) against the copied runtime closure, so a hook can't boot to ModuleNotFoundError fleet-wide from a missed file.
• r5-audit-gate (#25): risky-path changes require two independent adversarial audits before merge. Every change in this release went through it.

Notes

• No config changes required. Behavior change: supervisors will now receive a peer_idle on worker stops that were previously silent — this is the intended safety net.

Full log: v1.5.0..v1.6.0.

v1.5.0 — production-grade hardening (stable-install, boot gate, injection envelope, tool-only heartbeat)

claude-code-fleet-notify v1.5.0

The hardening release. Everything since v1.3.3 that made the fleet's notification + wake layer production-grade — every change merged through green CI gates with an adversarial audit by execution on the hook/stop surface.

New capabilities

• Wake-state packet delivery — stopped sessions wake with context (refs, memory, rules) injected through the hook path, not just a bare pointer.
• Prompt-injection hardening — inlined ref/memory/rule content in a wake packet is wrapped in a per-packet unforgeable nonce envelope with a data-only boundary, so file content can never be read as instructions. Verified inert by re-running real injection payloads.
• Receipt-validated handoffs — explicit handoff records + passive receipts, with delivery tracing (taey-trace).
• Tool-only liveness heartbeat — last_tool_activity is stamped only by tool hooks, closing the window where a daemon-injected prompt made an idle session look like it was working.
• Stop-policy enforcement via the orchestrator API.

Reliability fixes

• Stable runtime-root install + boot gate — hooks install to a stable location (never a movable checkout) and the installer proves every hook imports from its deployed location before writing any settings. Fixes the class of outage where a moved checkout could disable every tool on the machine.
• Symlink-safe CLI bootstrap for taey-handoff / taey-notify --handoff.

AI-native setup + honest docs

• readme-as-a-stranger CI gate — a fresh GitHub VM follows the README verbatim (install → daemon → round-trip) on every PR; if the docs lie, the check goes red. This is the guarantee an agent can stand the project up from zero.
• A live capability ledger (docs/CAPABILITIES.md) with a re-runnable observation command per row; README no longer hardcodes a version.

Notes

• Backed by adversarial review (grok + an independent gatekeeper) on every hook/stop-layer change, verified by execution.
• Single-machine tmux fleets; see the README scope section for what's in/out.

v1.3.3 — Horizon rollback-stop fix (STABLE)

5/5 Family unanimous ENDORSE achieved at this tag. Promote v1.3.3 as the verified-stable release.

What landed:

• Horizon amendment #5 — rollback() now explicitly stops the failed-target daemon BEFORE reverting symlink. start_notify_daemons.sh start is idempotent (refuses if pidfile shows alive), so without explicit stop a partially-started daemon on failed code persists while symlink flips back — leaving fleet in mixed state. Stop → flip → start sequence. JSONL audit line extended with daemon_stop_rc field for post-mortem visibility.

PR #10 merged. Companion fleet-orchestrator v1.3.3: Horizon amendments #2 (queue regression tests) + #3 (priority validation API guards) + Gaia 7th-site catch (plan_readiness.py).

v1.3.2 — atomic_deploy.sh rollback resilience (Cosmos full audit fold)

Folds Cosmos AMENDMENT 1: rollback() runs 'set +e' so a failed daemon restart can't prevent the JSONL audit log line from being written. Captures daemon restart return code into the JSONL payload (new daemon_restart_rc field).

Per Cosmos full audit (2026-05-31): 'a system that silently drops its audit log during a failure state violates our foundational transparency.'

PR #8 merged. Companion fleet-orchestrator v1.3.2 fold: coalesce(t.priority, 999999999) NULL guard across all task-priority sorts (Gaia AMENDMENT).

Superseded by v1.3.3.

v1.3.0 — Stop-Discipline Engine

What this ships

The fleet's notification hooks now route Stop events through a typed wake_type engine instead of the legacy peer_idle string-match path. Sessions that supervise projects with ready work get a WAKE_WITH_QUEUE carrying the next task; sessions with no ready work get WAKE_REASON_REQUIRED listing the project's user_stop_conditions to choose from; sessions with no supervised work or matched conditions get ALLOW_STOP. Engine errors that recur 3+ times in 60s trigger a bug_lock with ENGINE_ERROR wake.

Backwards-compat with v1.0.2 blocked_on is preserved: _resolve_blocked_on runs first and suppresses peer_idle when a worker stop is in response to a peer command.

What's new

• 4 typed wake_types: WAKE_WITH_QUEUE / WAKE_REASON_REQUIRED / ALLOW_STOP / ENGINE_ERROR
• 5 new CLIs: taey-stop-reason / taey-priority / taey-conditions / taey-pause / taey-queue
• File-flag feature gate: /home/mira/.taey/stage_b_enabled toggle for safe rollout (env-var fallback CF_STAGE_B_ENABLED=1 also works for daemon-spawned contexts)
• Atomic deploy script: scripts/atomic_deploy.sh with symlink-flip pattern + versioned directories + rollback trap + dual health gate (daemon [UP] + Redis ping()) + JSONL audit log
• Hot-deploy class developer guards: Makefile hooks-safe-edit target refuses operations on the live hooks-loading path; CI gate runs make test + gitleaks secret scan on every PR

Family audit lineage

• Stage B: 5/5 unanimous ENDORSE round 5 + Cosmos addendum (procedural BLOCK on packet construction — daemon.py + peer verdict citations were excluded from the packet though the prompt referenced them; resolved with verbatim inline addendum). Cosmos: "a map that skips coordinate verification because the system is tired is a map that lies." This is HMM.CANNOT_LIE_PROVENANCE working as designed.
• Stage B-Final: LIGHT scope round 1 (3 ENDORSE: Logos / Clarity / Cosmos + Horizon BLOCK on atomic_deploy.sh worktree-gitdir fragility + Gaia ENDORSE-WITH-AMENDMENTS on same plus a silent-success bootstrap bug) → round 2 re-verify (Horizon BLOCK LIFTED + Gaia ENDORSE non-blocking). Both Horizon's and Gaia's amendments folded into de868a6 + 707f744.

Rollback

Engine is rollback-able in under 2 seconds: rm /home/mira/.taey/stage_b_enabled. Hooks read the flag on every fire, so the rollback takes effect immediately, no daemon restart needed. Falls through to exact legacy v1.0.2 peer_idle path.

Deferred to v1.4.0

• Stage C UI panels for stop_reason management (5 CLIs cover the surface for now)
• Hook-version pinning (defense-in-depth atop CF gate)
• CI pre-deploy dry-run vs isolated Redis/Neo4j (current CI runs make test + gitleaks; service-container integration testing deferred)
• Recurring task templates (orchestrator supports task cycling but no built-in auto-instance-creation scheduler)

Known follow-ups

• Issue #5: atomic_deploy.sh worktree-gitdir fragility under multi-redeploy (works for first deploy + same-ref redeploy + different-ref swap; breaks under symlink-flip-then-redeploy because the deployed worktree's .git gitdir pointer references the source repo at creation time + the source path may move). Workaround: direct fresh-clone + manual symlink swap. Fix: replace git worktree add with git clone --depth 1 --branch <sha>.

Companion release

claude-code-fleet-orchestrator v1.3.0 — the Neo4j-backed orchestrator that this engine talks to.

🤖 Generated with Claude Code

v1.0.2

Blocked-On Stop-Hook Verification

Date: 2026-05-27

Scope

• [Observed] Verification exercised the live hooks/codex_stop.py entrypoint for the active tmux session conductor-codex.
• [Observed] The hook path resolved supervisor conductor.
• [Observed] Two real OrchTasks were created in the production tasks API and bound to the live session through the orchestrator dispatch wire:
  • blocked task: task-8af4e649
  • unblocked task: task-96cdbaa5

Blocked-On Suppression

• [Observed] task-8af4e649 was marked in_progress with blocked_on=family-round4-response.
• [Observed] Invoking the live Stop hook entrypoint produced stdout {} and stderr:

suppressed PEER_IDLE for conductor-codex: blocked_on=family-round4-response

• [Observed] taey:conductor:inbox had 0 messages before the blocked hook run and 0 messages after it.
• [Observed] No peer_idle message containing task-8af4e649 appeared in the supervisor inbox.

Inverse Check

• [Observed] task-96cdbaa5 was marked in_progress with blocked_on="".
• [Observed] Invoking the same live Stop hook entrypoint produced stdout {} and empty stderr.
• [Observed] taey:conductor:inbox had 0 messages before the unblocked hook run and 1 message after it.
• [Observed] The emitted message contained:
  • type=peer_idle
  • task_id=task-96cdbaa5
  • task_description="unblocked stop-hook notification verification"

Cleanup

• [Observed] The temporary current_task bindings for conductor-codex were cleared after verification.
• [Observed] The temporary supervisor inbox messages used for verification were removed after capture.

Unknowns

• [Unknown] This verification used the live Stop hook entrypoint directly rather than waiting for a human-driven CLI stop event. The code path, session identity detection, Redis state, and supervisor inbox were all production surfaces.

v1.0.1 — adopter-validation patches

Both treasurer and x-claude ran their next full cycle through the orchestrator stack and reported back. Gate question 'would you bet your next cycle on this without monitoring it manually?' — BOTH answered YES on this patch.

Fixes

Grok input-path (daemon.py)

For *-grok target sessions, daemon now injects FULL message body via tmux-send instead of the pointer summary.

Root cause (independently verified by treasurer + x-claude with 3-fact chains): grok-cli inherits Claude Code's hooks via ~/.claude/settings.json but does NOT honor the additionalContext field from hookSpecificOutput JSON the way Claude Code does. So prompt_activity.py drained the inbox + cleared idle correctly, but grok only ever saw the [NOTIFY] pointer text — never the full task body.

New build_grok_full_body() peeks the inbox + concatenates message bodies (capped at 6KB; truncates with redis-key pointer for the pathological case). Hook still drains on grok's submit (idempotent — bodies already delivered inline). Pointer pattern unchanged for Claude Code / codex / gemini.

Tmux-send grok submit sequence

Per-CLI detection via session-name suffix. Grok sequence: Ctrl-U + Ctrl-K (pre-clear) → text → plain Enter. NO Escape (toggles modes in grok-Build modal TUI). NO CSI-u Enter (also toggles modes in input-ready state). Claude/codex/gemini sequence unchanged. Pre-clear added to BOTH — prevents re-injection accumulation if prior submit didn't land.

taey-ack import fix

sys.path.insert now searches standard installed locations (/usr/local/lib/claude-code-fleet-notify, /home/mira/claude-code-fleet-notify, script-relative). Treasurer reported globally-broken ModuleNotFoundError: 'identity'; verified fixed.

Makefile install

make install now copies scripts/tmux-send AND scripts/start_notify_daemons.sh to $(PREFIX)/bin (both were missing).

Live verified

• Treasurer cycle: 8 tasks, 3 dispatches (1 gemini + 2 grok), all peer_idle structured wire validated, plan_readiness UNBLOCK wake fired exactly once. Grok produced real cannot-lie catches.
• x-claude cycle: dispatched cycle4-validation-grok-claim-gate-v2, grok read full body, executed real source-traceability work, wrote 5889-byte verdict file, record_outcome=done.

Pairs with

claude-code-fleet-orchestrator v1.0.1 (config robustness, state-drift grace, dispatch record_outcome footer).

v1.0.0 — first stable release

Marks production maturity. No code changes from v0.2.3 — bumps the semver signaling per Perplexity DR competitive analysis which flagged 0.x as the #1 HIGH-impact adoption headwind.

What's in v1.0.0 (cumulative since v0.1.0)

• Universal Stop+notify primitive across Claude Code, codex, gemini, grok via a single shared state machine + per-CLI hook variants.
• Lua compare-and-swap done-clear keyed on (worker, task_id) preventing silent task-loss races.
• Self-describing peer_idle wire with structured fields: task_id, task_description, task_supervisor, task_started_at, outcome, outcome_details, optional state_file + state_file_sha.
• Multi-CLI installer (install-hooks.sh --all) handling Claude Code + codex + gemini config formats; Grok inherits Claude Code via ~/.claude/settings.json.
• Complete scope marker (terminal-native hookable REPLs only; IDE-embedded agents and many-to-many distributed topologies explicitly out of scope).
• Lifecycle hooks for all four event types (PreToolUse / PostToolUse / Stop / UserPromptSubmit; gemini equivalents BeforeTool / AfterTool / BeforeAgent / AfterAgent).
• Local daemon for tmux pointer injection on idle sessions with pending inbox.

Family validation

Design consultations + 8-finding code audit + 5/5 system-integration sign-off (Gaia / Logos / Cosmos / Horizon / Clarity). Every Tier-1 closed before v1.0.0.

Live validation

Continuous operation on the Mira fleet running treasurer + conductor + x-claude workloads. Real-fleet dispatch verification 2026-05-26 through the structured peer_idle wire.

Semver promise

From v1.x: backwards-compatible bumps only. v2.x reserved for any breaking peer_idle envelope change.

Pairs with

claude-code-fleet-orchestrator v1.0.0 (companion bump same day).