Supervision Loop — Detailed Reference

Supervision Loop — Detailed Reference

The supervision loop runs BETWEEN each flow completion in silver-feature and silver-ui. It checks exit conditions, evaluates composition changes, detects stall, advances, and reports progress.

SL-1: Exit Condition Check (D-07.1)

Verify the flow's exit condition was met (per docs/composable-flows-contracts.md). If the exit condition is NOT met:

⚠ FLOW {name} exit condition not met: {condition description}
Options:
  A. Retry FLOW {name}
  B. Skip with reason (document in WORKFLOW.md)
  C. Insert FLOW DEBUG before next flow

SL-2: Composition Evaluation (D-07.2)

Re-evaluate context for dynamic insertion triggers:

• Execution failed → insert FLOW DEBUG before next flow (per D-11):
  • Record in WORKFLOW.md Dynamic Insertions table: | After {name} | FLOW DEBUG | Execution failed: {reason} | {timestamp} |
• UI files discovered in SUMMARY.md → insert FLOW DESIGN CONTRACT if not already in composition (per D-11, D-12):
  • Check SUMMARY.md for *.tsx, *.css, *.html, or design/ references
  • Record in WORKFLOW.md Dynamic Insertions table: | After {name} | FLOW DESIGN CONTRACT | UI files discovered | {timestamp} |

SL-3: Anti-Stall Check (D-07.3)

Run 4-tier anti-stall detection:

Tier 1 — Progress-based (D-16): If no WORKFLOW.md flow advancement in 10 minutes of execution wall-clock time, display:

⚠ STALL DETECTED: No flow advancement in 10 min. Continue? [Y/debug/skip]

Tier 2 — Permission-stall (D-17): If blocked waiting for user input >5 min in autonomous mode, auto-select recommended option (the first/default option) and log to WORKFLOW.md Autonomous Decisions table:

| {ISO timestamp} | Auto-selected option A for {decision} | Permission-stall: >5min wait in autonomous mode |

Tier 3 — Context exhaustion (D-18): Monitor context window usage:

• If context >80%: display Context compaction recommendation: context window at ~80%. Consider summarizing context before continuing.
• If context >90%: display Context exhaustion imminent. Summarize context before continuing. then run host-supported context compaction

Tier 4 — Heartbeat sentinel (D-19): Each flow invocation writes a heartbeat timestamp to WORKFLOW.md (Last-flow: and Last-beat: fields). If heartbeat gap >15 minutes, display:

⚠ HEARTBEAT GAP: FLOW {name} may have stalled. Options: [retry/skip/debug]

Heartbeat timestamps use ISO 8601 format (e.g., 2026-04-15T10:30:00Z).

SL-4: Advance (D-07.4)

Hook-driven (Wave 0.6): flow-advance.sh advances the orchestrator queue on each recorded flow-atom skill completion and emits the next flow. Do not inline agent logic to pick the next flow — invoke the skill named in the hook message.

Move to the next flow in the composition chain when the hook does not emit a next-flow message.

SL-5: Progress Report (D-09)

Display progress after each flow:

FLOW {current}/{total}: {name} ✓ | Context: ~{percent}% | Remaining: {list of remaining flows}

SL-6: WORKFLOW.md Update (D-10)

Write flow status and timestamp to WORKFLOW.md Flow Log table:

| {#} | FLOW {name} | complete | {artifacts produced} | ✓ |

Also update heartbeat fields:

Last-flow: {N}
Last-beat: {ISO timestamp}