Changelog

v0.3.x

Test reports: Centralized test reports in .dual-agent/test-reports/ with environment info. New ralph-lisa test-report command.

Watcher v5: Decoupled message delivery from response verification — fixes message flooding bug where watcher sent 14+ duplicate messages to working agents. Per-round send cap (max 2), capture-pane based idle detection (no pipe-pane dependency), pipe-pane cross-reference self-heal, passive post-send monitoring.
Mandatory test execution: [PLAN] must include test plan (test command + coverage scope). [CODE]/[FIX] Test Results must include exit code or pass/fail count. Explicit Skipped: with justification accepted. Lisa is required to re-run tests during review.
Escalation timing: Extended defaults from 2m/5m/10m to 5m/15m/30m. Configurable via RL_ESCALATION_L1, RL_ESCALATION_L2, RL_ESCALATION_L3 environment variables.
UX wording: Replaced "STOP immediately" / "MUST STOP" with "wait for feedback" across all role templates and command files.
Subagent guidance: Role templates now suggest using subagents for long-running tasks to avoid blocking the collaboration loop.
Smoke tests: 8 end-to-end CLI workflow scenarios with isolated test directories.

ralph-lisa stop command: Graceful shutdown of auto mode — stops watcher, sends /exit to agent panes, tears down tmux session. Supports --force for immediate kill and --no-archive to skip log archival.
Watcher v4: Round-based change detection fixes mutual deadlock where both agents stop saying they're waiting for each other. Uses monotonically increasing round number instead of turn value, so double-flip (A→B→A during long delivery) is always detected.
Cooldown bypass on new rounds: Cooldown timer no longer suppresses notifications for genuinely new rounds — only same-round re-delivery is throttled.
Consensus suppression with round boundary: Consensus detection now tracks which round it was detected at, preventing stale consensus from blocking notifications after next-step.
Crash recovery: Watcher state (SEEN_ROUND, ACKED_ROUND, DELIVERY_PENDING) persisted to .watcher_state file. On unexpected exit, state is preserved for wrapper restart replay. Graceful stop clears the file.
Escalation state machine: Multi-level stuck-agent detection (L1 REMINDER at 2min, L2 slash command at 5min, L3 user notification at 10min). Context limit detection jumps directly to L3. Delivery failure does not advance escalation level.

update-task command: Change task direction mid-session without restarting. Appends to task.md so history is preserved. Task context is auto-injected into submissions and watcher trigger messages.
Round 1 mandatory [PLAN]: Ralph's first submission must be [PLAN], giving Lisa a chance to verify understanding before coding begins.
Goal Guardian: Lisa now reads task.md before every review and checks for direction drift. Catching misalignment early is prioritized over code-level review.
Factual verification: Lisa must provide file:line evidence when claiming something is "missing" or "not implemented".
Policy layer: Configurable submission quality checks with warn/block modes.
Watcher v3: Fire-and-forget triggering, 30s cooldown, checkpoint system (RL_CHECKPOINT_ROUNDS), auto-restart on crash, configurable log threshold (RL_LOG_MAX_MB), heartbeat file.
Deadlock detection: After 8 consecutive [NEEDS_WORK] rounds, watcher auto-pauses for user intervention via scope-update or force-turn.
Minimal init: ralph-lisa init --minimal creates only session state (zero project files).
doctor command: Verify all dependencies with ralph-lisa doctor.

Fixed watcher deadlock: both agents simultaneously waiting for each other due to turn value comparison missing double-flip during long delivery.
Fixed consensus suppression blocking notifications after round change.
Fixed cooldown timer suppressing legitimate new-turn notifications within 30s window.
Fixed crash recovery: watcher state no longer deleted on unexpected exit (only on graceful stop).

Fixed case pattern escaping in generated watcher.sh — JS template literals silently stripped backslashes from case patterns, causing the watcher to crash-loop on every startup in auto mode.
Fixed check-next-step consensus logic to match step command behavior.
Fixed test isolation: neutralize tmux environment variables in test subprocesses.
Hardened watcher send-keys delivery for TUI agent compatibility.

Sharing the failures matters as much as the results:

Agent crashes have no auto-recovery. Once an agent crashes (possibly from long context or system resource exhaustion), the loop stops and you must manually restart. No self-healing yet.
State desync between agents. Early versions had Lisa going rogue — writing code herself instead of reviewing, causing state confusion. Much improved now, but the lesson stands.
Without domain judgment, the loop is useless. Two AIs will happily agree on a bad design. This is not autonomous development — it is structured AI-assisted development. The human arbiter isn't optional.
Git discipline is non-negotiable. Small commits, clear messages, commit often. When things go wrong (and they will), your only safety net is being able to git reset to a known good state.