Phase Gate Sequence

This skill enacts the Algorithms-Before-Tokens axiom from MANIFESTO.md: the 6-step gate is a deterministic algorithm encoded once and shared across all executive agents, eliminating parallel re-derivation and enforcing consistent phase discipline fleet-wide.

Beliefs & Context

• Governing axiom: Axiom 2 — Algorithms Before Tokens — deterministic procedure over repeated instruction
• GitHub issue: #79 — Skills as Decision Codifiers
• Formal FSM spec: data/phase-gate-fsm.yml — machine-readable state specification for this gate loop (states: INIT, PHASE_RUNNING, GATE_CHECK, COMPACT_CHECK, COMMIT, CLOSED)
• Agents that use this skill: Executive Orchestrator, Executive Researcher, Executive Fleet, Executive Docs
• Foundation docs:
  • AGENTS.md — compaction-aware writing, commit discipline, Programmatic-First
  • executive-orchestrator.agent.md — canonical per-phase sequence (lines 169–196) and context window alert protocol (lines 198–242)

Workflow

Run this sequence after every ## Phase N Output write, before delegating the next domain phase.

Step 0: Loop Audit (Pre-Phase)

Before any complex phase execution, run the orchestration loop detector:

uv run python scripts/detect_orchestration_loop.py \
  --task "<current-phase-name>" \
  --scratchpad ".tmp/$(git branch --show-current | tr '/' '-')/$(date +%Y-%m-%d).md"

• If loop_detected: false → proceed to Step 1
• If loop_detected: true → do not execute. Write to scratchpad: ## Loop Detected — [task] — iteration [N]; awaiting user direction. Surface to user: "I detected that I attempted this task [N] iterations ago. What should I do differently?"

Encoding point: Governed by AGENTS.md § Guardrails. Script: scripts/detect_orchestration_loop.py. Research basis: docs/research/orchestrator-autopilot-failure.md § Recommendation 4.

Step 1 — Prune

If the scratchpad exceeds 2000 lines:

uv run python scripts/prune_scratchpad.py

Step 1.5 — Pre-Delegation Rate-Limit Gate (Sprint 18+)

Before delegating the next phase, check the rate-limit budget gate (see rate-limit-resilience SKILL, issue #325):

# Get current token budget (from tracking or API)
CURRENT_BUDGET=75000  # example
OPERATION='delegation'  # or 'phase_boundary', 'fetch_source', etc.
PROVIDER='claude'

GATE_RESULT=$(uv run python scripts/rate_limit_gate.py "$CURRENT_BUDGET" "$OPERATION" --provider "$PROVIDER" --audit-log)

if echo "$GATE_RESULT" | grep -q '"safe": true'; then
    # Safe: proceed with delegation
    echo "Gate APPROVED: safe to proceed"
else
    # Blocked: defer or sleep
    SLEEP_SEC=$(echo "$GATE_RESULT" | grep -o '"recommended_sleep_sec": [0-9]*' | cut -d: -f2)
    echo "Gate BLOCKED: Rate-limited. Recommended sleep: ${SLEEP_SEC}s"
    echo "Defer this delegation to next session or sleep and retry"
    # Option 1: Write to scratchpad and stop
    # Option 2: Sleep and retry (caution: uses token budget)
fi

Integration: This check prevents cascading rate-limit failures mid-phase. If blocked:

• Log the gate decision to scratchpad under ## Rate-Limit Gate Output
• Defer the delegation to the next session, OR
• Sleep for recommended duration and monitor for consecutive failures (circuit-breaker)

Note: This is new infrastructure from Phase 0 (Sprint 18, issue #325). Phase 1 research (docs/research/ai-cognitive-load.md, issue #315) validates this gate: token-heavy workflows with >4 handoffs exceed working memory capacity; rate-limit gates reduce decision surface and human error. Integration point: AGENTS.md § Rate-Limit Resilience Throughout MANIFESTO Axioms shows how this gate operationalizes Algorithms-Before-Tokens (#319 trendslop research), Local-Compute-First (#317 vendor lock-in), and Endogenous-First (#315 cognitive load) axioms.

All orchestrators must check before high-cost delegations (e.g., research scout with 3+ parallel scouts).

Step 2 — Write Pre-Compact Checkpoint

Append ## Pre-Compact Checkpoint to the scratchpad with:

• What is complete (include commit SHAs)
• What is next
• Any open questions or blockers

Step 3 — Commit In-Progress Changes (Orchestrator)

Orchestrator commits in-progress changes via terminal:

git add -A && git commit -m "chore: pre-compact checkpoint — Phase N complete"

Step 4 — Pre-Review Grep Sweep

Scans the entire .github/ directory (all agent and skill files) for known heading-contract violations and erroneous patterns — not scoped to changed files only, since stale patterns can exist anywhere.

if grep -r "Phase N Review Output\|Fetch-before-check" .github/; then
  echo "ERROR: known pattern violations found — fix before requesting review"
elif [ $? -eq 1 ]; then
  echo "grep sweep clean"
else
  echo "ERROR: grep failed during sweep — investigate before requesting review"
fi

Fix any matches before invoking Review.

Step 5 — Review Gate

Invoke the Review agent with the changed file list and scratchpad location. Append verdict to the scratchpad under ## Review Output. Do not advance to the next phase until APPROVED.

Step 6 — Compact Recommendation

If the completed phase was a long research, synthesis, or multi-file editing delegation, recommend /compact before the next delegation. After any compaction event: re-read the scratchpad and workplan from disk before continuing.

Context Window Alert — Trigger Conditions

Pause all delegation immediately when any of the following is true:

• Compaction has occurred in this session, OR
• The user signals compaction is becoming frequent, OR
• You detect you are reconstructing already-explored context (re-reading the same file, re-running the same search)

When triggered:

1. Write ## Context Window Checkpoint to the scratchpad: active phase + status, committed vs. in-progress deliverables (with SHAs), last agent delegated + ≤100-token return summary, single next concrete step, open blockers.
2. Commit and push all in-progress changes.
3. Present the session handoff prompt template from executive-orchestrator.agent.md lines 215–236 — fill in bracketed fields from the scratchpad.
4. Do not proceed until the user resumes via the handoff prompt.

Completion Criteria

The phase gate is correctly applied for a given phase when:

• ## Pre-Compact Checkpoint exists in the scratchpad for every completed domain phase.
• ## Review Output with APPROVED verdict exists before the next domain phase was delegated.
• All phase deliverables are committed at gate time — no in-progress changes left dangling.
• Grep sweep ran clean (or violations were fixed before review was requested).

Exit condition: The loop terminates when ## Session Summary has been written and uv run python scripts/prune_scratchpad.py --force has run — see the session-management skill for the full session close sequence.