Runtime context (preprocessor snapshot):

• Active workflows: !find tmp -maxdepth 1 -name '.rune-*-*.json' -exec grep -l '"running"' {} + 2>/dev/null | wc -l | tr -d ' '
• Current branch: !git branch --show-current 2>/dev/null || echo "unknown"

/rune:debug — ACH-Based Parallel Debugging

Implements Analysis of Competing Hypotheses (ACH) methodology for multi-agent debugging. Instead of sequential hypothesis testing (systematic-debugging), this spawns parallel investigators — each assigned ONE hypothesis to confirm or falsify with evidence.

Load skills: systematic-debugging, rune-orchestration, context-weaving, team-sdk

When to Use

<!-- v3.x: defaults baked from former talisman.misc; see references/v3-defaults.md -->

Configuration (v3.x baked-in defaults)

In v3.x there is no talisman.yml user config layer — these debug values are inlined at the consumer call sites below:

See references/v3-defaults.md under ## misc for the canonical source-of-truth.

Phase 0: TRIAGE (Lead Agent)

Goal: Reproduce, classify, and generate competing hypotheses.

Step 0.1 — Parse Input

bugDescription = $ARGUMENTS

If $ARGUMENTS is empty, use AskUserQuestion to get:

• Bug description or failing test command
• Error output (if available)
• When it last worked (if known)

Step 0.2 — Reproduce the Bug

Run the failing test or reproduce the error:

testOutput = Bash("{test_command}")

If reproduction fails:

• Ask user for reproduction steps via AskUserQuestion
• If still cannot reproduce: report and exit

Record:

• First error: Exact error message at file:line
• Deterministic: Run 3x — same result each time?
• Recent changes: git log --oneline -10 -- {affected_files}

Step 0.3 — Classify Failure Category

Use trigger heuristics from ach-methodology.md:

1. Check CONCURRENCY first (intermittent failure?)
2. Check REGRESSION (recent commits to affected files?)
3. Check INTEGRATION (cross-module stack trace?)
4. Check ENVIRONMENT (env-specific error messages?)
5. Check DATA/FIXTURE (data-dependent, non-deterministic?)
6. Default: LOGIC ERROR

Record primary category and note any secondary matches.

Step 0.4 — Generate Hypotheses

Using templates from hypothesis-templates.md:

1. Generate 3-5 hypotheses from the primary category
2. Include at least 1 hypothesis from a different category (guard against misclassification)
3. Each hypothesis must be testable and falsifiable
4. Assign hypothesis IDs: {H-PREFIX}-{NNN} (e.g., H-REG-001)

Graceful degradation: If <2 hypotheses can be generated (bug is too simple or too unclear), fall back to single-agent systematic-debugging methodology. Do NOT spawn a team for 1 hypothesis.

Step 0.5 — Read Config

// v3.x: defaults baked from former talisman.misc; see references/v3-defaults.md
const maxInvestigators = 4
const timeoutMs = 420000
const investigatorModel = "sonnet"
const maxReTriageRounds = 1

Cap hypotheses at maxInvestigators.

Phase 1: INVESTIGATE (Parallel Agents)

Step 1.0 — Workflow Lock (writer)

CWD="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
source "${CWD}/plugins/rune/scripts/lib/workflow-lock.sh"
conflicts=$(rune_check_conflicts "writer")
if echo "$conflicts" | grep -q "CONFLICT"; then
  AskUserQuestion({ question: "Active workflow conflict:\n${conflicts}\nProceed anyway?" })
fi
rune_acquire_lock "debug" "writer"

Step 1.1 — Create Team

teamName = "rune-debug-{timestamp}"
TeamCreate({ team_name: teamName })

Fallback: If TeamCreate fails, fall back to single-agent systematic-debugging.

Step 1.2 — Create Investigation Tasks

For each hypothesis, create a task:

TaskCreate({
  subject: "Investigate {hypothesis_id}: {one-line summary}",
  description: `
    ## Assignment

    Investigate this ONE hypothesis. Gather confirming AND falsifying evidence.

    **Hypothesis ID**: {hypothesis_id}
    **Hypothesis**: {full hypothesis statement}
    **Category**: {category}

    ## Bug Context

    **Description**: {bugDescription}
    **Error output**: {testOutput}
    **Affected files**: {file_list}

    ## Evidence Standards

    Classify each evidence item by tier:
    - DIRECT (1.0): Uniquely produces/eliminates failure
    - CORRELATIONAL (0.6): Associated but alternate explanations exist
    - TESTIMONIAL (0.3): Reasoning without direct observation
    - ABSENCE (0.8/0.2): Expected evidence not found (0.8 exhaustive, 0.2 shallow)

    ## Output

    Write evidence report to: tmp/debug/{teamName}/{hypothesis_id}.md
    Use the format from your agent instructions.
  `,
  activeForm: "Investigating {hypothesis_id}"
})

Step 1.3 — MCP-First Investigator Discovery (v1.171.0+)

Before spawning investigators, query agent-search MCP for investigation agents:

// MCP-first investigator discovery — enables user-defined investigators
// (e.g., "perf-profiler" for performance-specific debugging)
let investigatorType = "hypothesis-investigator"  // default
try {
  const candidates = agent_search({
    query: "hypothesis investigation debugging root cause analysis evidence",
    phase: "goldmask",
    category: "investigation",
    limit: 5
  })
  Bash("mkdir -p tmp/.rune-signals && touch tmp/.rune-signals/.agent-search-called")

  if (candidates?.results?.length > 0) {
    // User-defined investigator can supplement default for specific debug contexts
    const userInvestigator = candidates.results.find(c =>
      (c.source === "user" || c.source === "project") &&
      c.name !== "hypothesis-investigator"
    )
    if (userInvestigator) investigatorType = userInvestigator.name
  }
} catch (e) {
  // MCP unavailable — use default hypothesis-investigator
}

Summon Investigators

For each hypothesis, spawn an investigator agent (all investigators use general-purpose subagent type — agent specialization comes from the prompt, not from the subagent type):

Agent({
  subagent_type: "general-purpose",
  team_name: teamName,
  name: "investigator-{N}",
  model: resolveModelForAgent("hypothesis-investigator", talisman),  // Cost tier mapping (../../references/cost-tier-mapping.md)
  prompt: "You are assigned hypothesis {hypothesis_id}. Claim your task from the task list, investigate, and report findings to tmp/debug/{teamName}/{hypothesis_id}.md"
})

Step 1.4 — Monitor Progress

Use polling loop (TaskList every cycle; see roundtable-circle/references/monitor-utility.md):

pollIntervalMs = 30000
maxIterations = ceil(timeoutMs / pollIntervalMs)

for iteration in 1..maxIterations:
  TaskList()  // MUST call TaskList every cycle (POLL-001)
  count completed tasks
  if all tasks completed: break
  if stale (no progress for 3 cycles): warn and continue
  Bash("sleep 30", { run_in_background: true })

Step 1.5 — Collect Reports

Read all evidence reports from tmp/debug/{teamName}/:

for each hypothesis_id:
  report = Read("tmp/debug/{teamName}/{hypothesis_id}.md")
  parse: verdict, confidence, evidence[], cross_signals

Handle missing reports (investigator timeout): proceed with available reports, note gap.

Phase 2: ARBITRATE (Lead Agent)

Execute the deterministic arbitration algorithm from ach-methodology.md.

Step 2.1 — Compute Scores

For each hypothesis with a report:

WES(H) = sum(supporting × tier_weight) - sum(refuting × tier_weight)
penalty = 0.4 × count(DIRECT cross-refutations)
FCS(H) = clamp((WES - penalty) × confidence_raw, 0, 1)

Step 2.2 — Rank and Threshold

1. Rank by FCS descending
2. Exclude REFUTED verdicts
3. Check thresholds:

• All FCS < 0.35: Re-triage (go to Step 2.4)
• FCS gap < 0.1 between top two: Tiebreaker (go to Step 2.3)
• Clear winner: Accept (go to Step 2.5)

Step 2.3 — Tiebreaker

Apply tiebreaker rules in order (see ach-methodology.md):

1. DIRECT evidence count
2. Absence of DIRECT refutation
3. Echo corroboration
4. Disproof test specificity
5. Compound hypothesis declaration

Step 2.4 — Re-Triage (if needed)

reTriageCount += 1
if reTriageCount > maxReTriageRounds:
  AskUserQuestion("All hypotheses scored below threshold. Here are the findings: {summary}. Can you provide additional context?")
  // Use user input to generate new hypotheses or exit
else:
  // Extract cross-hypothesis signals from all investigators
  // Generate new hypotheses
  // Return to Phase 1 (new investigation round)

Step 2.5 — Emit Verdict

Write verdict to tmp/debug/{teamName}/verdict.md:

# Debug Verdict — {teamName}

## Primary Hypothesis

**ID**: {winner_id}
**Statement**: {hypothesis}
**Confidence**: {HIGH|MEDIUM|LOW} (FCS: {score})
**Category**: {category}

## Evidence Summary

### Supporting (top 3 by weight)
1. [{evidence_id}] {description} — `{file:line}` (DIRECT, 1.0)
2. ...

### Refuting
- {refuting evidence if any}

## Refuted Alternatives

| Hypothesis | FCS | Reason |
|-----------|-----|--------|
| {H-XXX-NNN} | {score} | {key refuting evidence} |

## Cross-Hypothesis Signals

{signals that emerged during investigation}

## Recommended Fix

{Specific fix approach based on the winning hypothesis}

## Defense-in-Depth Layers

Based on category {category}, apply:
{layers from ach-methodology.md Defense-in-Depth Mapping}

Phase 3: FIX (Lead Agent)

Step 3.1 — Apply Fix

Based on the verdict, implement the fix:

• For REGRESSION: Revert or correct the regressed change
• For LOGIC: Fix the identified logic error
• For ENVIRONMENT: Add environment validation or documentation
• For DATA: Fix data source or add validation
• For INTEGRATION: Fix the boundary contract
• For CONCURRENCY: Add synchronization or fix ordering

Step 3.2 — Verify Fix

Run the original failing test/reproduction:

Bash("{original_test_command}")

If fix fails:

• Check if secondary hypothesis might be the actual cause
• If compound bug suspected, address both causes
• If stuck, escalate to user

Step 3.3 — Defense-in-Depth

Apply defensive layers per the failure category mapping in ach-methodology.md. Reference defense-in-depth.md for layer implementation details.

Step 3.4 — Echo Persistence (DEPRECATED)

v3.0.0-alpha.1: persistent memory layer (rune-echoes skill) was removed. The echo_on_verdict flag is retained for forward compatibility but is a no-op. A future version may reintroduce a lighter-weight verdict log.

Phase 4: CLEANUP

Standard 5-component team cleanup (QUAL-012): dynamic member discovery (fallback: investigator-{1..N}), shutdown broadcast, grace period, retry-with-backoff TeamDelete (4 attempts), process kill + filesystem fallback gated on !cleanupTeamDeleteSucceeded. Releases workflow lock. Then presents final summary (bug, hypothesis, fix, defense layers, rejected alternatives).

See phase-4-cleanup.md for the full cleanup pseudocode.

Arc Phase 7.7 Integration

When triggered by arc (test-failure-analyst returned LOW confidence):

1. Skip AskUserQuestion — use the test output and failure context from arc
2. Write verdict to tmp/debug/{teamName}/verdict.md for arc consumption
3. Do NOT apply fix directly — return verdict to arc for mend phase to handle
4. Set exit signal for arc dispatcher: write tmp/.rune-signals/{arc-team}/debug-complete.signal

Graceful Degradation