@accelerationguy/accel 1.0.0

package/modules/radar/src/core/workflows/phase-4-adversarial-review.md

@@ -0,0 +1,165 @@
+<purpose>
+Executes the adversarial review by invoking the Devil's Advocate agent to challenge all prior findings — attacking confidence, exposing blind spots, and surfacing what was missed across the complete analytical record from Phases 2-3.
+</purpose>
+
+<phase_context>
+Phase: 4 — Adversarial Review ("Fresh Eyes")
+Prior phase output: Phase 2 domain findings (.radar/findings/ from 8 agents), Phase 3 synthesis findings (staff-engineer, reality-gap-analyst), all disagreement records (.radar/disagreements/*.md)
+Agents invoked: devils-advocate (single agent, solo session)
+Output: Devil's Advocate critique (.radar/findings/devils-advocate.md), new disagreement records (.radar/disagreements/*.md), updated .radar/STATE.md
+</phase_context>
+
+<required_input>
+@.radar/STATE.md
+@.radar/MANIFEST.md
+@.radar/scope.md
+@.radar/findings/*.md (ALL finding files from Phases 2-3)
+@.radar/disagreements/*.md (ALL existing disagreement records)
+@~/.claude/radar/core/agents/devils-advocate.md
+@~/.claude/radar/core/personas/devils-advocate.md
+@~/.claude/radar/schemas/finding.md
+@~/.claude/radar/schemas/disagreement.md
+@~/.claude/radar/schemas/confidence.md
+@~/.claude/radar/rules/epistemic-hygiene.md
+@~/.claude/radar/rules/disagreement-protocol.md
+@~/.claude/radar/rules/agent-boundaries.md
+@~/.claude/radar/core/workflows/session-handoff.md
+@~/.claude/radar/core/workflows/disagreement-resolution.md
+</required_input>
+
+<process>
+
+<step name="validate_prerequisites" priority="first">
+1. Verify .radar/STATE.md shows Phase 3 complete:
+   a. Check phase_3_complete: true
+   b. Check .radar/findings/ contains Phase 2 AND Phase 3 agent finding files
+2. If Phase 3 is not complete:
+   a. Halt with error: "Phase 3 (Cross-Domain Synthesis) not complete. Run phase-3-cross-domain workflow first."
+   b. Do not proceed.
+3. Inventory the complete analytical record:
+   a. Count total findings across all agents
+   b. Count total disagreements (resolved and unresolved)
+   c. Identify which domains have coverage and which have gaps
+4. Update .radar/STATE.md:
+   a. current_phase: 4
+   b. phase_status: in_progress
+   c. agents_remaining: [devils-advocate]
+</step>
+
+<step name="assemble_complete_record" priority="blocking">
+1. Load ALL finding files from .radar/findings/:
+   a. Phase 2 agents: architect, data-engineer, security-engineer, compliance-officer, senior-app-engineer, sre, performance-engineer, test-engineer
+   b. Phase 3 agents: staff-engineer, reality-gap-analyst
+   c. Note any agents that failed (gaps in record)
+2. Load ALL disagreement records from .radar/disagreements/:
+   a. Both resolved and unresolved disagreements
+   b. Include resolution rationale for resolved ones (the Devil's Advocate may challenge resolutions)
+3. Load .radar/STATE.md with full phase metrics:
+   a. Finding counts per agent and domain
+   b. Severity distribution
+   c. Coverage gaps
+   d. Disagreement resolution status
+4. Load .radar/scope.md for audit boundary awareness.
+5. Estimate total context size — the Devil's Advocate receives the LARGEST context of any agent session.
+</step>
+
+<step name="prepare_devils_advocate_context" priority="blocking">
+1. Load devils-advocate agent manifest from src/core/agents/devils-advocate.md.
+2. Resolve component references:
+   a. Persona: src/core/personas/devils-advocate.md
+   b. Domains: none (empty — meta-reasoner operates on analytical record, not raw signals)
+   c. Tools: none (empty — no tool signals, only findings and disagreements)
+   d. Schemas: finding, disagreement, confidence
+   e. Rules: epistemic-hygiene, disagreement-protocol, agent-boundaries
+3. Build context bundle:
+   a. Persona specification (first — adversarial identity is critical)
+   b. Rules (second — especially disagreement protocol)
+   c. Complete finding record (third — ALL findings from ALL agents)
+   d. All disagreement records (fourth — including resolved ones)
+   e. Audit state and metrics (fifth — coverage awareness)
+   f. Scope document (sixth — boundary awareness)
+4. If context exceeds session budget:
+   a. Never trim persona or rules
+   b. Prioritize: high-severity findings > medium > low
+   c. Prioritize: unresolved disagreements > resolved ones
+   d. Prioritize: findings from domains with highest signal density
+   e. Log any trimmed content in .radar/STATE.md
+</step>
+
+<step name="invoke_devils_advocate" priority="blocking">
+1. Invoke devils-advocate session with the complete analytical record:
+   a. Primary directive: INVALIDATE conclusions, not agree
+   b. Challenge high-confidence claims that lack diverse evidence
+   c. Attack clean narratives — look for suspiciously tidy conclusions
+   d. Surface areas with little disagreement — consensus may hide blind spots
+   e. Use inversion: "Under what conditions does this become unsafe?"
+   f. Seek asymmetric failure: "What failure would be disproportionately damaging?"
+2. Session produces:
+   a. Critique findings — each challenges a specific prior finding or pattern
+   b. New disagreement records — structured challenges to existing conclusions
+   c. Blind spot identification — areas the audit may have missed entirely
+   d. Confidence attacks — findings where confidence is higher than evidence warrants
+3. All output must conform to schemas:
+   a. Critiques as findings per src/schemas/finding.md
+   b. Challenges as disagreements per src/schemas/disagreement.md
+   c. Confidence assessments per src/schemas/confidence.md
+4. On session completion:
+   a. Invoke session-handoff workflow (src/core/workflows/session-handoff.md)
+   b. Update .radar/STATE.md: agents_completed += [devils-advocate]
+</step>
+
+<step name="process_adversarial_output" priority="blocking">
+1. For each new disagreement raised by the Devil's Advocate:
+   a. Validate against src/schemas/disagreement.md
+   b. Verify it references valid finding IDs from prior phases
+   c. Assign disagreement ID: D-{NNN} (continuing from previous sequence)
+   d. Set status: open
+   e. Set principal_response: pending (REQUIRED — will be filled in Phase 5)
+   f. Write to .radar/disagreements/{disagreement-id}.md
+2. For each critique finding:
+   a. Validate against src/schemas/finding.md
+   b. Write to .radar/findings/devils-advocate.md
+3. Queue disagreement-resolution workflow for ALL open disagreements:
+   a. This includes Phase 2 disagreements still unresolved
+   b. Plus all new Phase 4 disagreements
+   c. The Principal Engineer MUST respond to every disagreement before Phase 5 synthesis
+</step>
+
+<step name="update_state_phase_complete" priority="blocking">
+1. Update .radar/STATE.md:
+   a. current_phase: 4
+   b. phase_status: complete
+   c. phase_4_complete: true
+   d. agents_completed: [devils-advocate]
+   e. devils_advocate_critiques: N (number of critique findings)
+   f. new_disagreements_phase_4: N
+   g. total_disagreements: updated cumulative count
+   h. unresolved_disagreements: count of status: open
+   i. next_phase: 5 (ONLY if all disagreements have principal_response, otherwise disagreement resolution first)
+   j. timestamp: completion time
+2. Determine next action:
+   a. If unresolved disagreements exist: route to disagreement-resolution workflow BEFORE Phase 5
+   b. If all disagreements resolved: route to Phase 5
+   c. Record routing decision in STATE.md
+</step>
+
+</process>
+
+<output>
+Artifacts created:
+- .radar/findings/devils-advocate.md — Adversarial critique findings
+- .radar/disagreements/{disagreement-id}.md — New challenge records for each contested conclusion
+- .radar/STATE.md — Updated with Phase 4 completion, critique counts, and disagreement status
+
+All finding files conform to src/schemas/finding.md.
+All disagreement records conform to src/schemas/disagreement.md.
+</output>
+
+<error_handling>
+- **Phase 3 not complete:** Halt immediately. Display: "Phase 3 (Cross-Domain Synthesis) must complete before adversarial review. Run phase-3-cross-domain first." Do not proceed.
+- **Devil's Advocate session failure:** This is a CRITICAL failure — Phase 4 is not optional. The adversarial review is a core epistemic safeguard. Record failure in STATE.md. Present to user with strong recommendation to retry. Do not auto-skip to Phase 5.
+- **Context budget exceeded:** The Devil's Advocate session has the largest context need. If the complete analytical record doesn't fit: (a) summarize low-severity findings instead of including full text, (b) include full text only for high/critical severity, (c) always include ALL disagreement records in full, (d) log what was summarized.
+- **Devil's Advocate produces no disagreements:** This is suspicious — a thorough adversarial review almost always finds something to challenge. Flag as potential issue: (a) Agent may not have received sufficient adversarial persona activation, (b) Agent may have been too deferential. Record concern in STATE.md. Do not block Phase 5 but note the anomaly.
+- **Excessive disagreements (>30% of findings challenged):** Not an error — a high challenge rate may indicate genuine issues. Record the volume. Do not auto-filter. All challenges go to the Principal Engineer for response.
+- **Disagreement references invalid finding ID:** Reject the specific disagreement record. Log the invalid reference. Accept other valid disagreements from the same session.
+</error_handling>

package/modules/radar/src/core/workflows/phase-5-report.md

@@ -0,0 +1,189 @@
+<purpose>
+Executes the final synthesis and report generation by invoking the Principal Engineer to consume the complete analytical record — all findings, all disagreements, and the Devil's Advocate critique — producing the definitive Radar audit report with mandatory response to every disagreement.
+</purpose>
+
+<phase_context>
+Phase: 5 — Synthesis & Report Generation
+Prior phase output: Complete .radar/ analytical record — Phase 2 domain findings, Phase 3 synthesis findings, Phase 4 Devil's Advocate critique, all disagreement records with principal_response
+Agents invoked: principal-engineer (Phase 5 role — synthesis and report generation, not scope establishment)
+Output: Final Radar report sections in .radar/report/, completed .radar/STATE.md marking audit as complete
+</phase_context>
+
+<required_input>
+@.radar/STATE.md
+@.radar/MANIFEST.md
+@.radar/scope.md
+@.radar/threat-model.md
+@.radar/findings/*.md (ALL finding files from ALL phases)
+@.radar/disagreements/*.md (ALL disagreement records — must all have principal_response)
+@~/.claude/radar/core/agents/principal-engineer.md
+@~/.claude/radar/core/personas/principal-engineer.md
+@~/.claude/radar/schemas/finding.md
+@~/.claude/radar/schemas/disagreement.md
+@~/.claude/radar/schemas/confidence.md
+@~/.claude/radar/schemas/report-section.md
+@~/.claude/radar/rules/epistemic-hygiene.md
+@~/.claude/radar/rules/disagreement-protocol.md
+@~/.claude/radar/rules/agent-boundaries.md
+@~/.claude/radar/core/workflows/session-handoff.md
+</required_input>
+
+<process>
+
+<step name="validate_prerequisites" priority="first">
+1. Verify .radar/STATE.md shows Phase 4 complete:
+   a. Check phase_4_complete: true
+   b. Check .radar/findings/devils-advocate.md exists
+2. CRITICAL CHECK — Verify ALL disagreements have principal_response:
+   a. Read every file in .radar/disagreements/
+   b. For each disagreement record:
+      - Verify principal_response field is present AND non-empty
+      - Verify status is not "open" (must be: mitigated, accepted_risk, deferred, or out_of_scope)
+   c. If ANY disagreement lacks principal_response:
+      - HALT. Do not proceed to Phase 5.
+      - Display: "Cannot synthesize report: N disagreements lack Principal Engineer response."
+      - List the unresolved disagreement IDs
+      - Route to disagreement-resolution workflow first
+3. If all prerequisites met:
+   a. Update .radar/STATE.md: current_phase: 5, phase_status: in_progress
+</step>
+
+<step name="assemble_complete_record" priority="blocking">
+1. Load the complete analytical record:
+   a. Phase 0: .radar/scope.md, .radar/threat-model.md
+   b. Phase 1: .radar/signals/SUMMARY.md (signal summary only — full signals not needed for synthesis)
+   c. Phase 2: .radar/findings/ from all 8 domain agents
+   d. Phase 3: .radar/findings/staff-engineer.md, .radar/findings/reality-gap-analyst.md
+   e. Phase 4: .radar/findings/devils-advocate.md
+   f. All disagreements: .radar/disagreements/*.md (with resolutions)
+   g. .radar/STATE.md with full cumulative metrics
+2. Organize by synthesis priority:
+   a. Critical/high severity findings first
+   b. Findings with associated disagreements grouped together
+   c. Cross-domain patterns (findings referenced by multiple agents)
+   d. Devil's Advocate critiques paired with the findings they challenge
+3. Estimate context size — Phase 5 has the SECOND largest context need (after Phase 4).
+4. If context exceeds budget:
+   a. Include ALL high/critical findings in full
+   b. Summarize medium/low findings by domain
+   c. Include ALL disagreements in full (resolutions are critical for the report)
+   d. Include ALL Devil's Advocate critiques in full
+   e. Log trimming decisions
+</step>
+
+<step name="invoke_principal_engineer" priority="blocking">
+1. Load principal-engineer agent manifest from src/core/agents/principal-engineer.md.
+2. Resolve component references:
+   a. Persona: src/core/personas/principal-engineer.md
+   b. Domains: none (meta-reasoner — synthesizes across all domains)
+   c. Tools: none (operates on analytical record)
+   d. Schemas: finding, disagreement, report-section, confidence, signal
+   e. Rules: epistemic-hygiene, disagreement-protocol, agent-boundaries
+3. Invoke principal-engineer session with Phase 5 instructions:
+   a. Synthesize the complete analytical record into the final Radar report
+   b. For every disagreement: verify the resolution is sound and the rationale holds under the full record
+   c. Identify cross-domain patterns that individual agents couldn't see
+   d. Calibrate severity across domains — ensure consistent severity scale
+   e. Produce a narrative that a technical leader can act on
+4. Session produces the 7 report sections:
+   a. Executive Risk Summary — top-line risk assessment for leadership
+   b. Architecture Narrative — structural health story
+   c. Findings by Domain — severity-ranked findings per domain (all 14 domains)
+   d. Cross-Validation Notes — disagreements and their resolutions with rationale
+   e. Remediation Roadmap — prioritized action plan
+   f. Long-Term Structural Risks — risks that compound over time
+   g. "What Would Break First at 10x Scale" — growth stress analysis
+5. Each report section must conform to src/schemas/report-section.md.
+</step>
+
+<step name="validate_report_output" priority="blocking">
+1. Verify all 7 report sections were produced:
+   a. Check each section exists and is non-empty
+   b. Validate each against src/schemas/report-section.md
+2. Verify disagreement coverage:
+   a. Every disagreement ID that exists in .radar/disagreements/ must appear in the Cross-Validation Notes section
+   b. No disagreement may be silently omitted — this is the core anti-pattern "hiding disagreements in footnotes"
+3. Verify severity calibration:
+   a. No finding severity was silently changed without explanation
+   b. If the Principal Engineer adjusted severity, the rationale is documented
+4. Verify cross-domain coherence:
+   a. The Architecture Narrative is consistent with the Findings by Domain section
+   b. The Remediation Roadmap addresses the highest-severity findings
+   c. The Executive Summary accurately reflects the detailed findings
+5. If validation fails:
+   a. If session is still available: return with specific gaps
+   b. If session ended: flag incomplete sections in STATE.md
+</step>
+
+<step name="persist_report" priority="blocking">
+1. Write report sections to .radar/report/:
+   a. .radar/report/01-executive-summary.md
+   b. .radar/report/02-architecture-narrative.md
+   c. .radar/report/03-findings-by-domain.md
+   d. .radar/report/04-cross-validation-notes.md
+   e. .radar/report/05-remediation-roadmap.md
+   f. .radar/report/06-long-term-risks.md
+   g. .radar/report/07-scale-analysis.md
+2. Generate combined report:
+   a. Assemble all sections into .radar/report/Radar-REPORT.md
+   b. Include table of contents
+   c. Include audit metadata (scope, timeline, agent roster, tool versions)
+3. Invoke session-handoff workflow for final time:
+   a. Capture Principal Engineer session output
+   b. Persist any final findings to .radar/findings/principal-engineer.md (append to Phase 0 findings)
+   c. Update STATE.md
+4. Verify all report files written successfully.
+</step>
+
+<step name="finalize_audit" priority="blocking">
+1. Update .radar/STATE.md:
+   a. current_phase: 5
+   b. phase_status: complete
+   c. audit_status: complete
+   d. phase_5_complete: true
+   e. report_sections: 7
+   f. total_findings_all_phases: N
+   g. total_disagreements: N
+   h. disagreements_resolved: N (should equal total)
+   i. completion_timestamp: time
+2. Record final audit metrics:
+   a. Total duration (Phase 0 start to Phase 5 end)
+   b. Findings by severity: critical, high, medium, low, info
+   c. Domain coverage: which domains produced findings
+   d. Agent participation: which agents completed successfully
+   e. Disagreement resolution rate: 100% required
+3. Mark .radar/MANIFEST.md as final:
+   a. Add report hash for integrity verification
+   b. Lock manifest — no further modifications to this audit
+4. Present audit completion summary to user:
+   a. Report location: .radar/report/Radar-REPORT.md
+   b. Key statistics
+   c. If Transform pipeline is configured: note readiness for Layer B/C generation
+</step>
+
+</process>
+
+<output>
+Artifacts created:
+- .radar/report/01-executive-summary.md — Executive Risk Summary
+- .radar/report/02-architecture-narrative.md — Architecture health narrative
+- .radar/report/03-findings-by-domain.md — Severity-ranked findings per domain
+- .radar/report/04-cross-validation-notes.md — Disagreements and resolutions
+- .radar/report/05-remediation-roadmap.md — Prioritized action plan
+- .radar/report/06-long-term-risks.md — Compounding structural risks
+- .radar/report/07-scale-analysis.md — "What breaks at 10x scale" analysis
+- .radar/report/Radar-REPORT.md — Combined final report
+- .radar/findings/principal-engineer.md — Updated with Phase 5 findings (appended)
+- .radar/STATE.md — Finalized with audit_status: complete
+
+All report sections conform to src/schemas/report-section.md.
+</output>
+
+<error_handling>
+- **Phase 4 not complete:** Halt immediately. Display: "Phase 4 (Adversarial Review) must complete before synthesis. Run phase-4-adversarial-review first." Do not proceed.
+- **Unresolved disagreements:** CRITICAL BLOCK. Phase 5 MUST NOT proceed if any disagreement lacks principal_response. Route to disagreement-resolution workflow. This is non-negotiable — silence on disagreements destroys audit trust.
+- **Principal Engineer session failure:** This is a CRITICAL failure — the audit cannot produce a report without Phase 5. Record failure. Present to user with strong recommendation to retry. Do not produce a partial report automatically.
+- **Incomplete report (missing sections):** If fewer than 7 sections produced, flag which sections are missing. If session is available, request completion. If not, mark audit as "partial" in STATE.md and note which sections are absent.
+- **Context budget exceeded:** Phase 5 has large context needs. Trimming strategy: (a) keep ALL disagreements in full, (b) keep ALL critical/high findings in full, (c) summarize medium/low by domain, (d) keep Devil's Advocate critiques in full, (e) summarize signal summary. Never trim persona or rules.
+- **Report inconsistency detected:** If validation finds the Executive Summary contradicts detailed findings, or the Remediation Roadmap doesn't address highest-severity issues: flag specific inconsistencies. If session available, return for correction. If not, annotate the report with consistency warnings.
+</error_handling>

package/modules/radar/src/core/workflows/phase-checkpoint.md

@@ -0,0 +1,222 @@
+<purpose>
+Provides a structured checkpoint between Radar audit phases — displaying phase completion summary, cumulative progress, and offering the user a clear continue/pause/abort decision point. Invoked by the audit command after each phase workflow completes.
+</purpose>
+
+<phase_context>
+Phase: Utility — invoked between every phase transition during Core audit execution (0→1, 1→2, 2→3, 3→4, 4→5)
+Prior phase output: Completed phase findings, signals, and updated .radar/STATE.md
+Agents invoked: None — this workflow presents information and captures user intent
+Output: Updated .radar/STATE.md with checkpoint record, user decision (continue/pause/abort)
+</phase_context>
+
+<required_input>
+@.radar/STATE.md
+@.radar/MANIFEST.md
+</required_input>
+
+<process>
+
+<step name="capture_phase_completion" priority="first">
+1. Read .radar/STATE.md to identify which phase just completed.
+2. Extract from STATE.md:
+   - completed_phase: the phase number (0-5) that just finished
+   - phase_status: confirm it is "complete"
+   - agents_completed: list of agents that ran in this phase
+   - findings_produced: count of findings from this phase
+   - disagreements_produced: count of disagreements from this phase (if any)
+3. Count actual finding files in .radar/findings/ for this phase:
+   - Cross-reference agent IDs from completed phase against finding files
+   - Tally by severity: critical, high, medium, low, info
+4. Record phase completion timestamp in .radar/STATE.md if not already set.
+5. Increment session tracking counters in .radar/STATE.md:
+   - Increment Sessions count
+   - Update Last session timestamp
+   - Set Started timestamp if this is the first checkpoint
+</step>
+
+<step name="display_phase_summary" priority="blocking">
+Display the completed phase summary:
+
+```
+════════════════════════════════════════
+PHASE [N] COMPLETE — [Phase Name]
+════════════════════════════════════════
+
+Agents completed: [count] ([agent-id list])
+Findings produced: [count]
+  Critical: [N]  High: [N]  Medium: [N]  Low: [N]
+Disagreements: [N] detected
+
+────────────────────────────────────────
+CUMULATIVE PROGRESS
+────────────────────────────────────────
+
+[██████████░░░░░░░░░░] Phase [N] of 5 complete
+
+| Phase | Name                       | Status   | Findings |
+|-------|----------------------------|----------|----------|
+|   0   | Context & Threat Modeling  | complete | -        |
+|   1   | Signal Gathering           | complete | -        |
+|   2   | Deep Domain Audits         | [status] | [count]  |
+|   3   | Cross-Domain Synthesis     | [status] | [count]  |
+|   4   | Adversarial Review         | [status] | [count]  |
+|   5   | Synthesis & Report         | [status] | -        |
+
+Total findings so far: [N]
+Total disagreements: [N] (resolved: [N])
+Sessions used: [N]
+════════════════════════════════════════
+```
+
+Phase names reference:
+- Phase 0: Context & Threat Modeling
+- Phase 1: Automated Signal Gathering
+- Phase 2: Deep Domain Audits
+- Phase 3: Change Risk & Reality Gap
+- Phase 4: Adversarial Review
+- Phase 5: Synthesis & Report
+</step>
+
+<step name="preview_next_phase" priority="blocking">
+If completed_phase < 5 (more phases remain):
+
+1. Determine next phase number: completed_phase + 1
+2. Look up next phase details:
+
+   | Phase | Agents | Estimated Sessions | Description |
+   |-------|--------|--------------------|-------------|
+   | 1     | (tools) | 1 session        | Automated tool scanning — SonarQube, Semgrep, Trivy, etc. |
+   | 2     | 8      | 1-8 sessions      | Domain specialist agents analyze findings in parallel |
+   | 3     | 2      | 1-2 sessions      | Staff Engineer + Reality Gap Analyst (sequential) |
+   | 4     | 1      | 1 session          | Devil's Advocate challenges all findings |
+   | 5     | 1      | 1 session          | Principal Engineer synthesizes final report |
+
+3. Display preview:
+   ```
+   ────────────────────────────────────────
+   NEXT: Phase [N] — [Name]
+   ────────────────────────────────────────
+
+   [Description]
+   Agents: [count or "tools"]
+   Estimated sessions: [range]
+   Prerequisites: Phase [N-1] complete ✓
+   ────────────────────────────────────────
+   ```
+
+If completed_phase == 5 (Core audit complete):
+   ```
+   ────────────────────────────────────────
+   CORE AUDIT COMPLETE
+   ────────────────────────────────────────
+
+   All 6 diagnostic phases finished.
+   Total findings: [N]
+   Disagreements resolved: [N] of [N]
+
+   Next steps:
+   - /radar:report — generate the final audit report
+   - /radar:transform — start the remediation pipeline
+   ────────────────────────────────────────
+   ```
+</step>
+
+<step name="offer_checkpoint_options" priority="blocking">
+If completed_phase < 5 (more phases remain):
+
+Present exactly 3 options:
+```
+[1] Continue to Phase [N+1] (recommended)
+[2] Pause here — safe to close this session
+[3] Abort audit — preserves all work completed so far
+```
+
+Handle each response:
+
+**If [1] "continue":**
+- Update .radar/STATE.md Resume Info:
+  - Last action: "Phase [N] checkpoint — user chose continue"
+  - Next action: "Phase [N+1] — [Name]"
+- Return control to the audit command's phase loop to invoke the next phase workflow
+
+**If [2] "pause":**
+- Update .radar/STATE.md:
+  - Overall status: paused
+  - Resume Info:
+    - Last action: "Phase [N] checkpoint — user chose pause"
+    - Next action: "Resume with /radar:resume to start Phase [N+1]"
+- Display:
+  ```
+  ════════════════════════════════════════
+  AUDIT PAUSED
+  ════════════════════════════════════════
+
+  All progress saved to .radar/STATE.md.
+  [N] findings preserved across [N] phases.
+
+  To continue later:
+    /radar:resume
+
+  To check status anytime:
+    /radar:status
+  ════════════════════════════════════════
+  ```
+- Stop execution (do not proceed to next phase)
+
+**If [3] "abort":**
+- Update .radar/STATE.md:
+  - Overall status: paused
+  - Resume Info:
+    - Last action: "Phase [N] checkpoint — user chose abort"
+    - Next action: "Resume with /radar:resume or start fresh with /radar:audit"
+- Display:
+  ```
+  ════════════════════════════════════════
+  AUDIT STOPPED
+  ════════════════════════════════════════
+
+  All findings preserved. Nothing was deleted.
+  Completed phases: 0 through [N]
+  Findings: [count] total
+
+  To resume later:  /radar:resume
+  To start fresh:   /radar:audit (will offer to archive existing state)
+  ════════════════════════════════════════
+  ```
+- Stop execution
+
+If completed_phase == 5 (Core audit complete):
+- No continue/pause/abort — just display completion info and suggest next steps
+- Update .radar/STATE.md: overall status to "complete"
+</step>
+
+<step name="update_state" priority="required">
+Regardless of user choice, update .radar/STATE.md:
+
+1. Add to Checkpoint History section:
+   ```markdown
+   - Phase [N] ([Name]) — [timestamp] — [continue/pause/abort]
+   ```
+
+2. Update Session Tracking:
+   - Increment Sessions count
+   - Update Last session to current timestamp
+
+3. Update phase completion timestamp in the Phases table if not already set.
+</step>
+
+</process>
+
+<output>
+- .radar/STATE.md updated with checkpoint record, session tracking, and user decision
+- User informed of progress and given clear next-step guidance
+- If "continue": control returns to audit command for next phase
+- If "pause" or "abort": execution stops cleanly with resume instructions
+</output>
+
+<error_handling>
+- **STATE.md missing or unreadable:** Display error "Cannot read .radar/STATE.md — audit state may be corrupted. Run /radar:status to diagnose." Do not proceed.
+- **Phase not actually complete:** If STATE.md shows phase still active (agents remaining), do not display checkpoint. Warn: "Phase [N] has [X] agents remaining. Complete the phase before checkpointing."
+- **Session Tracking section missing:** If STATE.md was created before Phase 12 updates (no Session Tracking section), add it now with Sessions: 1, Last session: current timestamp, Started: current timestamp.
+- **Checkpoint History section missing:** If missing, create it with the current checkpoint as the first entry.
+</error_handling>