Fellow

Fellow is the system's empirical optimization engine, invoked exclusively by Mentor to determine which implementation of a skill, prompt, heuristic, or workflow actually performs best — not which one looks best on paper. It runs controlled experiments with a fixed benchmark and compute budget, establishes a fresh baseline before testing any variant, and returns the winning result with full mutation lineage so every promotion is traceable and reversible.

Interactive Menu

When invoked interactively, present a two-level menu. See references/interactive-menu.md for the full menu structure.

When to Use

Fellow is not user-invocable. It is called only by Mentor when:

• A skill's OKR performance has regressed
• A variant proposal needs empirical evaluation
• A prompt, heuristic, or workflow needs optimization
• Mentor needs to compare champion vs challenger implementations

For example, when Mentor proposes a new heuristic for skill promotion, Fellow runs a controlled experiment to compare it against the baseline.

When NOT to Use

• User-initiated requests — Fellow is Mentor-only
• Skill building or design — use Forge
• Pattern analysis — query Chronicle directly
• Web research — use Sift

Responsibility boundary

Fellow owns empirical experimentation: baseline establishment, variant generation, benchmark execution, metric extraction, and promotion decisions. This division exists because Mentor lacks the compute budget for controlled experiments, while Fellow lacks strategic direction.

Fellow does not own: deciding what to improve (Mentor), building skill packages (Forge), behavioral refinement (Praxis).

Mentor provides direction. Fellow provides empirical optimization.

Ontology Types

Fellow observes entity types during experiment execution (Concept/Idea, Thing/DigitalArtifact, Concept/Event). Fellow includes entity observations in journal outputs for Chronicle ingestion. See references/schemas.md for full details.

Invocation Guard

Fellow is not user-invocable. If triggered directly by a user prompt, respond: "Fellow is an internal engine invoked only by Mentor for benchmark experiments. For skill evaluation, use Mentor."

Inter-Skill Interfaces

Mentor → Fellow: Fellow reads ExperimentRequest files from Mentor's experiment-requests directory. Fellow → Mentor: Fellow writes CycleResult files to {agent_root}/commons/data/ocas-fellow/results/. See spec-ocas-interfaces.md for schemas.

Experiment Lifecycle

See references/schemas.md for the full experiment lifecycle (8 steps), baseline protocol, mutation engine, promotion rule, cycle output schema, run completion procedure, and failure handling.

Skillgrade evaluation engine: Fellow natively executes declarative eval.yaml suites (from a skill's references/evals/eval.yaml) inside isolated ocas-inception containers — challenger variants run with zero external side effects. Execution: run each eval task across N trials, apply deterministic + LLM-rubric graders, and return a CycleResult containing trial pass rates, token metrics, and per-grader breakdowns to Mentor (written to {agent_root}/commons/data/ocas-fellow/results/). These results feed Mentor's pass-rate promotion thresholds (≥0.85 / N≥5, N≥15 for core execution skills). See spec-ocas-skill-improvements.md §1.

Commands

• fellow.experiment.run — execute an experiment cycle from Mentor invocation payload
• fellow.experiment.status — current experiment state if in progress
• fellow.journal — write journal for the current run; called at end of every run
• fellow.update — pull latest from GitHub source; preserves journals and data

Recovery Behavior

This skill implements the recovery contract from spec-ocas-recovery.md.

• Evidence: Every experiment run writes evidence to {agent_root}/commons/data/ocas-fellow/evidence.jsonl, including no-op runs with mandatory not_activity_reason.
• Gap detection: On every wake, checks evidence log for most recent run. If gap exceeds 24h for update cron, logs gap_detected.
• Degraded mode: When Mentor or experiment harness unavailable, logs degraded: <dependency> and queues work for retry.
• Log compaction: Evidence logs older than 30 days (no-op) or 90 days (error/gap) compacted. Escalation records never auto-deleted. Last 7 days retained.

Storage layout

See references/schemas.md → Storage layout and Default config.json.

OKRs

Universal OKRs from spec-ocas-journal.md apply to all runs. See references/schemas.md → OKRs for the full skill_okrs definition.

Optional skill cooperation

• Mentor — sole invoker; provides experiment programs and approves promotions
• Chronicle — experiment lineage and entity observations written via journal signal payloads

Journal Outputs

Action Journal — every experiment cycle execution. Entity observations (Concept/Idea, Thing/DigitalArtifact, Concept/Event) may be included for lineage tracking. Each entity observation includes a user_relevance field (user, agent_only, unknown). See references/journal.md for full schema.

Initialization

On first invocation by Mentor, run fellow.init: create data directories, write default config, create empty JSONL files, register cron. See references/schemas.md for full initialization steps.

Background Tasks

# Task declared in SKILL.md frontmatter metadata.{platform}.cron

Self-update

fellow.update pulls the latest package from the source: URL in this file's frontmatter. Runs silently — no output unless the version changed or an error occurred. See references/schemas.md → Self-update procedure for the full step-by-step.

Visibility

public

Gotchas

See references/gotchas.md for all operational pitfalls including invocation guard, baseline failure, entity observation handling, promotion threshold, and Mentor/Fellow responsibility split.

Support File Map