alekspetrov/nav-loop

Navigator Loop Skill

Execute tasks iteratively until completion with structured signals, stagnation detection, and dual-condition exit gates.

Why This Exists

Traditional AI coding requires manual "keep going" prompts. Navigator Loop provides:

• Structured completion signals (NAVIGATOR_STATUS block)
• Dual-condition exit gate (heuristics + explicit signal)
• Stagnation detection (circuit breaker for stuck loops)
• Progress visibility (phases: INIT → RESEARCH → IMPL → VERIFY → COMPLETE)

Based on Ralph's autonomous loop innovations, adapted for Navigator's context-efficient architecture.

When to Invoke

Auto-invoke when:

• User says "run until done", "keep going until complete"
• User says "iterate until finished", "autonomous mode"
• User says "loop mode", "don't stop until done"
• Task document has loop_mode: true

DO NOT invoke if:

• Single-step task (no iteration needed)
• User says "just do this once"
• Already in loop mode (prevent nested loops)
• User explicitly disabled loop mode

Configuration

Loop mode settings in .agent/.nav-config.json:

{
  "loop_mode": {
    "enabled": false,
    "max_iterations": 5,
    "stagnation_threshold": 3,
    "exit_requires_explicit_signal": true,
    "show_status_block": true,
    "iteration_approval": "none",
    "periodic_interval": 3,
    "never_pause_on_stagnation": false,
    "stagnation_diversify_strategy": "combine"
  }
}

Core options:

• enabled: Default state for new tasks
• max_iterations: Hard cap to prevent infinite loops (1-20)
• stagnation_threshold: Same-state count before pause (2-5)
• exit_requires_explicit_signal: Require EXIT_SIGNAL alongside heuristics
• show_status_block: Render NAVIGATOR_STATUS each iteration

Autonomous / overnight options (v6.2.2+):

• iteration_approval: When to prompt the user for accept/reject between iterations.
  • "none" (default) — never prompt; loop runs uninterrupted
  • "strict" — prompt after every iteration
  • "periodic" — prompt every N iterations (where N = periodic_interval, default 3)
• periodic_interval: When iteration_approval == "periodic", the cadence of prompts. Default 3 (every 3rd iteration). Set higher for less frequent check-ins on long overnight runs (e.g., 5 or 10).
• never_pause_on_stagnation: If true, stagnation triggers auto-diversification instead of an AskUserQuestion pause. Required for true overnight runs. Inspired by karpathy/autoresearch's NEVER STOP directive.
• stagnation_diversify_strategy: Which recovery to attempt when never_pause_on_stagnation fires.
  • "combine" — combine previous near-misses / partially-met indicators
  • "radical" — try a substantially different approach (re-architect, swap library)
  • "reread" — re-read the in-scope task/system docs for missed signals

Safety guard: Setting never_pause_on_stagnation: true REQUIRES max_iterations to be set explicitly (the default of 5 is fine; the point is — no infinite default). Without a max, an autonomous loop can spin forever on a fundamentally broken task.

Execution Steps

Step 1: Initialize Loop State

Load configuration:

python3 functions/phase_detector.py --init

Initialize tracking variables:

iteration = 1
max_iterations = config.loop_mode.max_iterations or 5
stagnation_threshold = config.loop_mode.stagnation_threshold or 3
hash_history = []
phase = "INIT"

Display loop start:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
LOOP MODE ACTIVATED
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Task: {TASK_DESCRIPTION}
Max iterations: {max_iterations}
Stagnation threshold: {stagnation_threshold}

Starting iteration 1...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Step 2: Execute Iteration

Perform task work based on current phase:

| Phase | Actions | |-------|---------| | INIT | Load context, understand requirements | | RESEARCH | Explore codebase, find patterns | | IMPL | Write code, make changes | | VERIFY | Run tests, validate functionality, simplify code | | COMPLETE | All indicators met, ready to exit |

Track changes during iteration:

• Files read (for RESEARCH detection)
• Files changed (for IMPL detection)
• Tests run (for VERIFY detection)
• Commits made (for completion indicator)

Step 3: Generate Status Block

After each iteration, generate NAVIGATOR_STATUS:

python3 functions/status_generator.py \
  --phase "{phase}" \
  --iteration "{iteration}" \
  --max-iterations "{max_iterations}" \
  --indicators "{indicators_json}" \
  --state-hash "{current_hash}" \
  --prev-hash "{previous_hash}" \
  --stagnation-count "{stagnation_count}"

Display status block:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
NAVIGATOR_STATUS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Phase: {PHASE}
Iteration: {N}/{MAX}
Progress: {PERCENT}%

Completion Indicators:
  [{x or space}] Code changes committed
  [{x or space}] Tests passing
  [{x or space}] Code simplified
  [{x or space}] Documentation updated
  [{x or space}] Ticket closed
  [{x or space}] Marker created

Exit Conditions:
  Heuristics: {MET}/{TOTAL} (need 2+)
  EXIT_SIGNAL: {true/false}

State Hash: {HASH}
Previous Hash: {PREV_HASH}
Stagnation: {COUNT}/{THRESHOLD}

Next Action: {NEXT_ACTION}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Step 3.5: Per-Iteration Approval Gate (optional)

Skip this step if config.loop_mode.iteration_approval == "none" (default).

Run this step if the user wants oversight between iterations — for risky changes, learning the loop's behavior, or sanity-checking before a long run.

| Setting | Behavior | |---------|----------| | "none" | Never prompt. Loop continues to Step 4. | | "strict" | Prompt after every iteration. | | "periodic" | Prompt every Nth iteration. N defaults to 3; configurable via loop_mode.periodic_interval. |

When prompting, use AskUserQuestion immediately after the status block:

Question: "Accept iteration {N} and continue?"
Options:
  1. [Continue] - Iteration accepted, proceed to next
  2. [Adjust]   - Provide feedback, incorporate into next iteration
  3. [Abort]    - End loop, create partial-completion marker

Decision handling:

• Continue: Proceed to Step 4 normally.
• Adjust: Capture the user's feedback into a transient note, do NOT advance hash history (so the next iteration is judged as fresh progress), continue to Step 4.
• Abort: Jump to Step 8 (Cleanup) with status: "user_aborted".

This gate runs BEFORE stagnation detection so that a rejected iteration doesn't accidentally accumulate stagnation count.

Step 4: Check Stagnation

Calculate state hash:

python3 functions/stagnation_detector.py \
  --phase "{phase}" \
  --indicators "{indicators_json}" \
  --files-changed "{files_json}" \
  --history "{hash_history_json}"

If stagnation detected (same hash for N iterations), the response depends on never_pause_on_stagnation.

Default behavior (never_pause_on_stagnation: false)

Prompt the user:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
STAGNATION DETECTED
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Same state detected for {N} consecutive iterations.

Current State:
  Phase: {PHASE}
  Indicators: {MET}/{TOTAL}
  Last Action: {LAST_ACTION}

Possible causes:
1. Blocked by external dependency
2. Unclear requirements
3. Test failures preventing progress
4. Missing context or permissions

Options:
1. [Continue] - Try one more iteration
2. [Clarify] - Explain what's blocking
3. [Abort] - End loop, manual intervention

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Use AskUserQuestion for choice:

• Continue: Reset stagnation counter, continue loop
• Clarify: User explains blocker, incorporate and continue
• Abort: Exit loop with partial completion marker

Autonomous behavior (never_pause_on_stagnation: true)

The loop is running unattended (overnight, CI, etc.). Do NOT prompt — auto-diversify based on stagnation_diversify_strategy:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
STAGNATION → AUTO-DIVERSIFY ({STRATEGY})
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Same state for {N} iterations. Auto-recovery: {STRATEGY}
Stagnation counter reset.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

| Strategy | What to attempt next iteration | |------------|--------------------------------| | combine | Re-examine partially-met indicators; combine 2 near-miss approaches from previous iterations | | radical | Discard the current approach; try a substantially different design (different library, different architecture, different algorithm) | | reread | Re-read the in-scope task doc + relevant system docs; look for a missed signal or constraint |

After diversifying:

• Reset stagnation counter to 0
• Record the diversification in the iteration's notes (so the next status block reflects it)
• Continue to Step 5

Hard stop: Even in autonomous mode, the loop still terminates on max_iterations. If diversification has been triggered ≥3 times within a single run, escalate to the abort path (creates a loop-aborted marker and exits) — repeated diversification is itself a signal that the task is fundamentally stuck.

Inspired by karpathy/autoresearch's NEVER STOP directive: "If you run out of ideas, think harder — read papers, re-read in-scope files for new angles, try combining previous near-misses, try more radical architectural changes."

Step 5: Check Exit Conditions

Evaluate dual-condition gate:

python3 functions/exit_gate.py \
  --indicators "{indicators_json}" \
  --exit-signal "{exit_signal}" \
  --require-explicit "{config.exit_requires_explicit_signal}"

Exit conditions:

1. Heuristics: At least 2 completion indicators met
2. EXIT_SIGNAL: Explicit signal that task is complete

Completion indicators (mapped from autonomous protocol):

• code_committed: Changes committed to git
• tests_passing: Test suite passes (exit code 0)
• code_simplified: Code simplified for clarity (v5.4.0+)
• docs_updated: Documentation files changed
• ticket_closed: PM tool ticket marked done
• marker_created: Completion marker exists

Exit decision logic:

IF heuristics >= 2 AND exit_signal == true:
  → EXIT: Task complete
ELIF heuristics >= 2 AND exit_signal == false:
  → CONTINUE: Awaiting explicit completion signal
ELIF exit_signal == true AND heuristics < 2:
  → BLOCKED: Cannot exit with insufficient indicators
ELSE:
  → CONTINUE: More work needed

Step 6: Handle Max Iterations

If iteration >= max_iterations:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
MAX ITERATIONS REACHED
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Completed {MAX} iterations without full completion.

Current State:
  Phase: {PHASE}
  Indicators: {MET}/{TOTAL}
  EXIT_SIGNAL: {true/false}

Progress made:
- {PROGRESS_ITEM_1}
- {PROGRESS_ITEM_2}

Options:
1. [Extend] - Add 3 more iterations
2. [Complete] - Accept current state as done
3. [Abort] - Exit without completion

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Step 7: Complete Loop

When exit conditions met, emit the exit signal in JSON format and display completion:

{"v":2,"type":"exit","success":true,"reason":"All criteria met"}

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
LOOP COMPLETE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Task: {TASK_DESCRIPTION}
Iterations: {FINAL_COUNT}/{MAX}
Final Phase: COMPLETE

Completion Indicators:
  [x] Code changes committed
  [x] Tests passing
  [x] Code simplified
  [x] Documentation updated
  [ ] Ticket closed (skipped - no PM tool)
  [x] Marker created

Exit Conditions:
  Heuristics: 4/5 (passed)
  EXIT_SIGNAL: true (passed)

Summary:
- {KEY_CHANGE_1}
- {KEY_CHANGE_2}
- {KEY_CHANGE_3}

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Execute autonomous completion protocol:

1. Commit changes (if not already)
2. Archive task documentation
3. Close ticket (if PM configured)
4. Create completion marker (with loop state)
5. Suggest compact

---

Setting EXIT_SIGNAL

The EXIT_SIGNAL is set explicitly by Claude when:

• All primary task requirements are met
• Code is functional and tested
• No obvious remaining work

How to signal completion (v2 JSON format):

{"v":2,"type":"exit","success":true,"reason":"All requirements met"}

The JSON format in a pilot-signal code block ensures unambiguous detection by Pilot automation. The reason field should briefly describe why the task is complete.

Signal fields:

• v: Version (always 2)
• type: Signal type (always "exit" for completion)
• success: Whether task completed successfully (true/false)
• reason: Brief explanation of completion state

This explicit declaration prevents premature exits when heuristics are met but work remains.

---

Phase Detection

Phases auto-detected based on context:

def detect_phase(context):
    # COMPLETE: Exit conditions met
    if indicators_met >= 4 and exit_signal:
        return "COMPLETE"

    # VERIFY: Tests running or recently run
    if context.tests_running or context.test_exit_code is not None:
        return "VERIFY"

    # IMPL: Files being modified
    if context.files_changed:
        return "IMPL"

    # RESEARCH: Reading files, searching
    if context.files_read and not context.files_changed:
        return "RESEARCH"

    # INIT: Default starting state
    return "INIT"

---

Integration with Navigator

With Autonomous Completion

Loop mode enhances (not replaces) the autonomous protocol:

• Completion indicators map to autonomous steps
• EXIT_SIGNAL triggers autonomous completion
• Marker includes loop state for restoration

With nav-simplify (v5.4.0+)

Simplification runs during VERIFY phase:

• After tests pass, before committing
• Configurable via simplification.enabled in .nav-config.json
• Adds code_simplified completion indicator
• Skip if no code changes (docs-only tasks)

With nav-diagnose

Stagnation triggers nav-diagnose quality check:

• 3 same-state loops = potential quality issue
• nav-diagnose helps identify root cause
• Re-anchoring can resolve stuck loops

With nav-marker

Markers capture loop state:

• Current iteration and max
• Phase at time of marker
• State hash for continuity
• Completion indicators status

With ToM Features

Loop mode respects ToM configuration:

• Verification checkpoints still apply in VERIFY phase
• Profile preferences affect communication style
• Belief anchors can help clarify stuck states

---

Predefined Functions

functions/status_generator.py

Generates formatted NAVIGATOR_STATUS block.

functions/exit_gate.py

Evaluates dual-condition exit (heuristics + explicit signal).

functions/stagnation_detector.py

Calculates state hash and detects consecutive same-states.

functions/phase_detector.py

Auto-detects current task phase from context.

---

Error Handling

Config not found:

Loop mode config not found in .nav-config.json.
Using defaults: max_iterations=5, stagnation_threshold=3

Function execution fails:

• Fall back to manual evaluation
• Log error but don't interrupt loop
• Continue with best-effort phase detection

User aborts mid-loop:

• Create partial completion marker
• Document progress made
• List remaining work

---

Success Criteria

Loop mode succeeds when:

• [ ] Task completes within max_iterations
• [ ] No stagnation pauses (or resolved quickly)
• [ ] EXIT_SIGNAL + heuristics both satisfied
• [ ] Completion marker includes loop state
• [ ] User sees clear progress each iteration

---

Examples

Example 1: Simple Feature

User: "Run until done: add isPrime function with tests"

Iteration 1 (INIT → RESEARCH):
  - Read existing math utils
  - Found test patterns

Iteration 2 (IMPL):
  - Created isPrime function
  - Created test file

Iteration 3 (VERIFY):
  - Ran tests: PASS
  - Committed changes

{"v":2,"type":"exit","success":true,"reason":"isPrime function implemented and tests passing"}

→ Loop complete in 3 iterations

Example 2: Stagnation Recovery

User: "Run until done: fix authentication bug"

Iteration 1-3 (IMPL):
  - Same changes attempted
  - Tests still failing
  - State hash unchanged

→ STAGNATION DETECTED

User: "The test needs a mock for the auth service"

Iteration 4 (IMPL):
  - Added mock
  - Tests pass

{"v":2,"type":"exit","success":true,"reason":"Auth bug fixed with mock service"}

→ Loop complete in 4 iterations

---

Limitations

Cannot handle:

• External blockers (waiting for API, permissions)
• Subjective completion criteria ("make it look nice")
• Tasks requiring human judgment mid-loop

Should not use for:

• Quick fixes (single iteration sufficient)
• Exploratory work (no clear completion state)
• Tasks with security implications (need human review)

---

This skill provides Ralph-style "run until done" capability while maintaining Navigator's context efficiency and ToM integration.