Refs: PR #426
Features:
- fcctl-core to terraphim_firecracker adapter
- Sub-500ms VM allocation (267ms measured)
- ULID-based VM ID enforcement
- Full trait implementation with error preservation
- 119 tests passing

Validation: All acceptance criteria met

…erTier enum

Add provider, fallback_provider, fallback_model, and provider_tier fields
to AgentDefinition for subscription-based model routing (ADR-002, ADR-003).
Add ProviderTier enum (Quick/Deep/Implementation/Oracle) with per-tier
timeout values. Add opencode CLI support in spawner arg inference.

All new fields are Optional with serde(default) for backward compatibility.

Fixes #28
Refs #29

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Add provider allowlist and banned_providers fields to OrchestratorConfig.
Default banned list includes "opencode" (Zen pay-per-use proxy, ADR-002).
Add validate_provider() method to spawner AgentConfig that rejects any
model string starting with a banned prefix, while correctly allowing
opencode-go/ (subscription) and kimi-for-coding/ (subscription).

Fixes #31

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Add persona_name, persona_symbol, persona_vibe, meta_cortex_connections,
and skill_chain fields to AgentDefinition for the four-layer identity
stack (Persona/Role/SFIA/Skills). All fields Optional/default for
backward compatibility. See ADR-004.

Fixes #32

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Add SpawnRequest struct and spawn_with_fallback() method that:
- Validates banned providers before spawn (ADR-002)
- Uses ProviderTier timeout values (Quick=30s, Deep=60s, Impl=120s, Oracle=300s)
- Retries with fallback provider/model on primary failure
- Integrates per-provider circuit breakers (3 failures = open 5 min)
- Returns SpawnerError on both primary and fallback failure

SpawnRequest avoids circular dependency with terraphim_orchestrator
by mirroring needed AgentDefinition fields in the spawner crate.

Fixes #30

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Add persona_name, persona_symbol, persona_vibe, and meta_cortex_connections
to SpawnRequest. The build_persona_prefix() function generates a markdown
identity block that is prepended to the agent task prompt when persona
fields are configured. Agents without persona config are unaffected.

Fixes #33

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Add SkillChainRegistry with 31 terraphim-skills and 16 zestic-skills.
Provides validate_chain() to verify skill chains and validate_skill_chains()
on OrchestratorConfig to validate all agents. Backward compatible -- agents
with empty skill_chain pass validation.

Fixes #34

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Add OpenCodeEvent struct with parsing for opencode run --format json
output events (step_start, text, tool_use, step_finish, result).
Includes text_content(), total_tokens(), parse_line(), and parse_lines()
helper methods for structured output consumption.

Fixes #37

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Add SkillResolver in skill_resolver.rs that maps skill chain names to actual
skill file paths from the terraphim-skills repository.

Features:
- SkillResolver with registry of terraphim skills (security-audit, code-review,
  session-search, local-knowledge, git-safety-guard, devops, disciplined-research,
  architecture, disciplined-design, requirements-traceability, testing,
  acceptance-testing, documentation, md-book, implementation, rust-development,
  visual-testing, quality-gate)
- resolve_skill_chain() method that takes Vec<skill_name> and returns
  resolved skill metadata (name, description, applicable_to, paths, source)
- Validation of skill chains with proper error handling
- Comprehensive test coverage including valid resolution, missing skill errors,
  and empty chain handling
- Exported from crate as SkillResolver, SkillSource, ResolvedSkill,
  SkillResolutionError

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…h Refs #36

Extend SkillResolver to support zestic-engineering-skills from zestic-ai/6d-prompts
repository alongside existing terraphim-skills.

Features:
- Add SkillSource enum (Terraphim, Zestic) to distinguish skill origins
- Initialize zestic skills registry with 12 skills: quality-oversight,
  responsible-ai, insight-synthesis, perspective-investigation, product-vision,
  wardley-mapping, business-scenario-design, rust-mastery, cross-platform,
  frontend, via-negativa-analysis, strategy-execution
- Each resolved skill includes its source metadata
- SkillChainRegistry validation accepts skills from both sources
- Comprehensive tests for mixed skill chains (terraphim + zestic together)

Tests added:
- test_resolver_has_zestic_skills
- test_resolve_zestic_skill
- test_resolve_mixed_skill_chain
- test_validate_mixed_skill_chain_valid
- test_validate_only_zestic_skills
- test_mixed_chain_with_invalid_skills
- test_zestic_skill_source_in_resolved
- test_all_skill_names_includes_zestic
- test_zestic_skill_structure
- test_resolver_custom_paths_for_zestic

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

…efs #38

…routing Refs #39

Replace all legacy codex CLI references with opencode + subscription providers.
Add persona, skill_chain, provider/model/fallback fields to all agents.
New agents: compliance-watchdog, drift-detector, spec-validator, test-guardian,
documentation-generator, implementation-swarm, compound-review.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

SSH non-interactive shells (systemd) do not include ~/.bun/bin or
~/.local/bin in PATH. Use full paths for opencode and claude binaries.
Also fix TOML escape in --since quote.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Creates crates/terraphim_judge_evaluator with:
- SimpleAgent struct wrapping terraphim_router for KG lookups
- KgMatch struct for representing term matches
- lookup_terms() method using Aho-Corasick automata
- enrich_prompt() for appending KG context to judge prompts
- Comprehensive tests for all functionality

Also adds automation/judge/model-mapping.json configuration file.

Implements JudgeModelRouter for tier-based LLM model selection:
- TierConfig struct for provider+model pairs
- ModelMappingConfig for complete configuration
- JudgeModelRouter with methods:
  * from_config(path) - load from JSON file
  * resolve_tier(tier) - get provider/model for a tier
  * resolve_profile(profile) - get tier sequence for a profile
- Supports tiers: quick, deep, tiebreaker, oracle
- Supports profiles: default, thorough, critical, exhaustive
- Comprehensive tests for all functionality
- Updates automation/judge/model-mapping.json with profile definitions

Implements JudgeAgent as a supervised agent combining:
- SimpleAgent for Knowledge Graph context enrichment
- JudgeModelRouter for tier-based model selection
- SupervisedAgent trait from terraphim_agent_supervisor

JudgeVerdict struct with:
- verdict: final evaluation result (PASS/FAIL/NEEDS_REVIEW)
- scores: BTreeMap of detailed category scores
- judge_tier: which tier produced the verdict
- judge_cli: command used for evaluation
- latency_ms: evaluation duration

Evaluation pipeline:
1. Load file content
2. Enrich with KG context (optional)
3. Select model tier based on profile
4. Parse verdict from CLI output (JSON or text format)

Includes comprehensive tests for:
- Verdict parsing and manipulation
- Supervised agent lifecycle
- Profile-based tier selection
- System message handling

- Add stagger_delay_ms config field (default: 5000ms)
- Insert stagger delay between Safety agent spawns in run()
- Add random jitter (0 to stagger_delay_ms) for Core agent cron spawns
- Add tests for stagger delay configuration

Refs #16

- Add ReviewRequest struct with from_agent, to_agent, artifact_path, review_type
- Add ReviewPair config for defining (producer, reviewer) pairs
- Add review_queue: Vec<ReviewRequest> to orchestrator state
- Add submit_review_request(), review_queue(), process_review_queue() methods
- Add check_review_trigger() to automatically queue reviews on agent completion
- Add tests for review queue operations and config loading

Refs #17

- Add drift_detection module with DriftDetector and DriftReport
- Load strategic goals from plans/ directory markdown files
- Check drift every N ticks (configurable via drift_detection.check_interval_ticks)
- Calculate drift score by comparing agent outputs against strategic goals
- DriftReport includes agent, drift_score, and explanation
- Log warnings when drift_score exceeds threshold (default: 0.6)
- Add DriftDetectionConfig to OrchestratorConfig
- Add tests for drift detection functionality

Refs #18

- Add session_rotation module with SessionRotationManager and AgentSession
- Add SessionRotationConfig with max_sessions_before_rotation (default: 10)
- Track completions_since_rotation and completed_sessions per agent
- on_agent_completion() records completion and triggers rotation at threshold
- Rotation creates new session ID and clears accumulated context
- Add comprehensive tests for session rotation

Refs #19

- Add convergence_detector module with ConvergenceDetector and ConvergenceSignal
- Add ConvergenceConfig with threshold (default: 0.95) and consecutive_threshold (default: 3)
- Calculate output similarity using Jaccard index on word sets
- Detect convergence after N consecutive similar outputs
- ConvergenceSignal includes agent, similarity, and consecutive_count
- Reset on divergence to handle changing outputs
- Add comprehensive tests for convergence detection

Refs #20

…efs #23

…iven mode

Issue #8: Extend orchestrator config for issue-driven mode
- Add WorkflowConfig with mode (time_only/issue_only/dual), poll_interval_secs, max_concurrent_tasks
- Add TrackerConfig with tracker_type (gitea/linear), url, token_env_var, owner, repo
- Add ConcurrencyConfig with max_parallel_agents, queue_depth, starvation_timeout_secs
- All optional in orchestrator.toml with backward compatible defaults (time_only mode is default)
- Tests: parse config with and without workflow section, defaults applied

Issue #9: Unified dispatcher
- Create crates/terraphim_orchestrator/src/dispatcher.rs
- DispatchTask enum: TimeTask(agent_name, schedule), IssueTask(agent_name, issue_id, priority)
- DispatchQueue: priority queue backed by BinaryHeap with round-robin fairness
- ConcurrencyController: semaphore-based with starvation timeout
- Methods: submit(task), next() -> Option<DispatchTask>, active_count(), is_full()
- Fairness: alternates between time and issue tasks at equal priority
- Tests: submit/dequeue, priority ordering, concurrency limits, fairness

Issue #10: Issue mode controller
- Create crates/terraphim_orchestrator/src/issue_mode.rs
- IssueMode struct using terraphim_tracker::GiteaTracker to poll for issues
- Poll loop every poll_interval_secs fetching ready issues via PageRank sorting
- Filter out blocked issues and already-running tasks
- Map issues to agents based on labels ([ADF] -> implementation-swarm) or title patterns
- Submit IssueTask to DispatchQueue
- Tests: poll cycle, issue-to-agent mapping, priority calculation, blocked issue filtering

Issue #11: Time mode refactor
- Refactor crates/terraphim_orchestrator/src/scheduler.rs
- Extract TimeMode struct wrapping existing cron scheduler
- TimeMode submits TimeTask to DispatchQueue instead of spawning directly (when configured)
- Maintain backward compatibility: legacy mode spawns directly if no WorkflowConfig
- Tests: TimeMode submits to queue, legacy mode still works

Refs #8 #9 #10 #11

- Add ModeCoordinator that manages TimeMode and IssueMode
- Implement unified shutdown: signal modes, drain queue, wait for active tasks
- Extend spawner integration to dispatch tasks from queue to agents
- Add stall detection: log warning when queue exceeds threshold
- Add comprehensive tests for dual mode, shutdown coordination, and stall detection

Refs #12

Create comprehensive E2E test suite in tests/e2e_tests.rs:
- test_dual_mode_operation: verify both time and issue tasks processed
- test_time_mode_only: legacy config compatibility
- test_issue_mode_only: issue-only config verification
- test_fairness_under_load: no starvation between task types
- test_graceful_shutdown: clean termination with queue draining
- test_stall_detection: warning logged when queue exceeds threshold
- Additional tests for concurrency limits, prioritization, and backward compatibility

Uses mock tracker and avoids real API calls.

Refs #13

- Run cargo test --workspace (orchestrator tests pass)
- Fix all warnings in orchestrator crate
- Create CHANGELOG.md entry for v1.9.0 release
- Verify backward compatibility with test_backward_compatibility
- Build release binary: cargo build --release -p terraphim_orchestrator
- All 142 tests pass in orchestrator crate

Release includes:
- Dual mode orchestrator with TimeMode and IssueMode
- Unified dispatch queue with fairness
- Stall detection and graceful shutdown
- Symphony compatibility layer
- Comprehensive E2E test suite
- Migration documentation

Refs #15

- Fix needless borrows for generic args in workspace crate
- Apply consistent code formatting across modified files