Mythos-Inspired Reliability Harness — orthogonal scouts, self-testing lead, clean-checkout verifier, deterministic Done Gate.
A configuration bundle for OpenCode that applies observable, evidence-based reasoning patterns derived from published frontier-model research. Not a model swap, not a jailbreak, not a 1:1 emulation.
Star this repo if you want to follow the empirical validation work.
Fable & Mythos for OpenCode CLI is a configuration package for OpenCode that applies observable, evidence-based reasoning patterns (multi-option exploration, multi-criteria evaluation, auditability, strategic reasonableness) through a multi-agent reliability harness with a deterministic Done Gate.
This is not a model swap. This is not a jailbreak. This is not a 1:1 emulation of any other model. It is a behavioral priming framework + sub-agent orchestration protocol that plugs into OpenCode's native agents/, skills/, and AGENTS.md (instructions) mechanisms.
The hypothesis: Independent, evidence-based verification improves reliability. Every non-trivial coding task is processed through specialized sub-agents — orthogonal read-only scouts, a self-testing lead engineer, and a clean-checkout verifier — ending in a deterministic, machine-observed Done Gate. Empirical validation against a GLM-5.2 baseline is planned, not yet measured.
| Dimension | Status | Note |
|---|---|---|
| Reasoning depth | Unrated | Multi-option + multi-criteria evaluation per agent |
| Output reliability | Unrated | Clean-checkout verification + Done Gate |
| Anti-hallucination | Unrated | Hypothesis only; benchmark pending |
| Ease of install | Unrated | Copy 3 directories + managed-block merge |
| OpenCode integration | Unrated | Native agents/skills/AGENTS.md; compatible with OmO |
This package brings a reliability harness inspired by published frontier-model reasoning research to OpenCode CLI. OpenCode natively supports custom agents (via agents/), skills (via skills/), and global instructions (via AGENTS.md). This package uses all three for a structured multi-agent verification protocol.
- Custom agents — auto-discovered from
agents/<name>.md(frontmatter:name,description,mode,tools). - Skills —
skills/<name>/SKILL.mdwith frontmatter; loaded on demand via the Skill tool. - Global instructions —
AGENTS.mdloaded via OpenCode's auto-discovery (and optionally via"instructions": ["..."]). - Per-agent tool control —
tools: [...]list in agent frontmatter (least privilege).
Mythos-inspired OpenCode · OpenCode CLI subagents · OpenCode custom agents · OpenCode multi-agent reasoning · OpenCode reliability harness · OpenCode verification protocol · Done Gate · clean-checkout verifier
MAP = Multi-Agent Verification Protocol. Fires automatically on non-trivial coding tasks, with dynamic routing by risk tier (see core/routing.md).
| Risk tier | Agents engaged |
|---|---|
| trivial | Main agent alone |
| normal | Main agent + clean-checkout verifier |
| complex | Orthogonal read-only scouts (parallel) → Lead (implements + self-tests) → Verifier |
| critical | complex + Adversary in isolated worktree |
The legacy 3× parallel identical-thinking fleet is kept as an optional fallback but is not the default — three orthogonal scouts produce real diversity (codebase / spec / verification) rather than stylistic variants of the same assumption.
flowchart LR
T[Non-trivial<br/>Coding Task] --> P0
subgraph P0[Phase 0 — Orthogonal read-only scouts]
direction TB
S1[reliability-scout<br/>call graph, affected files,<br/>conventions, existing tests]
S2[reliability-spec-critic<br/>acceptance contract,<br/>ambiguities, scope]
S3[reliability-test-designer<br/>own worktree: repro,<br/>regression, edge cases]
end
P0 -->|structured outputs| L
subgraph P1[Phase 1 — Implement + Self-test]
L[reliability-lead<br/>implements, runs tests,<br/>diagnoses, repairs]
end
L -->|frozen patch| V
subgraph P2[Phase 2 — Clean-checkout verification]
direction TB
V[reliability-verifier<br/>fresh worktree, 9-point check]
A[reliability-adversary<br/>only at risk_tier=critical]
end
P2 --> G{Done Gate<br/>machine-observed}
G -->|VERIFIED| OUT[Final delivery<br/>with evidence]
G -->|PARTIALLY_VERIFIED / BLOCKED| L
| # | Agent file | Role | Tools |
|---|---|---|---|
| 0 (legacy) | agents/mythos-singleshot-thinking-intelligence.md |
Optional parallel thinking. Emits thinking pass only. | read, grep, glob |
| 1 (legacy) | agents/mythos-executor.md |
Builds artifact + self-tests. | read, edit, write, bash, grep, glob |
| 2 (legacy) | agents/mythos-verifier.md |
9-point verification on clean checkout. | read, bash, grep, glob |
| 3 (legacy) | agents/mythos-adversary.md |
Red-team, 12 attack vectors. | read, bash, grep, glob |
| 4 (legacy) | agents/mythos-synthesizer.md |
Aggregates findings (recommendation only; Done Gate decides). | read, grep, glob |
| 5 | agents/reliability-scout.md |
Call graph, affected files, conventions, existing tests. | read, grep, glob |
| 6 | agents/reliability-spec-critic.md |
Acceptance contract, ambiguities, scope, invariants. | read, grep, glob |
| 7 | agents/reliability-test-designer.md |
Repro, regression, edge cases (own worktree). | read, edit, write, bash, grep, glob |
| 8 | agents/reliability-lead.md |
Implements + self-tests (Lead Engineer standard). | read, edit, write, bash, grep, glob |
| 9 | agents/reliability-verifier.md |
Clean-checkout verification, 9-point check. | read, bash, grep, glob |
| 10 | agents/reliability-adversary.md |
Only at risk_tier=critical: fuzzing, race/security probes. |
read, bash, grep, glob |
| Task type | Behavior |
|---|---|
| Coding task with substance | Full harness fires automatically (risk-tier routed) |
| Trivial edit (typo, 1-line fix) | Skipped |
| Pure info questions | Skipped |
| Ambiguous | Fires |
# Clone the package
git clone https://github.com/emco1234/fable-mythos-opencode.git ~/fable-mythos-opencode
# Copy components into OpenCode config
cp -r ~/fable-mythos-opencode/agents/* ~/.config/opencode/agents/
cp -r ~/fable-mythos-opencode/skills/fable-mythos-modus ~/.config/opencode/skills/
# Merge AGENTS.md using the idempotent managed-block installer
# (see INSTALLATION.md Step 3 for the full idempotent script)Then edit ~/.config/opencode/opencode.json to set permissions. See INSTALLATION.md for the full merge walkthrough (including the idempotent managed-block installer).
opencode
# Inside OpenCode TUI, the agents should be auto-discovered
# Test by invoking: "Use reliability-lead to refactor this function"Full walkthrough: INSTALLATION.md
fable-mythos-opencode/
├── README.md ← You are here
├── AGENTS.md ← Global rules (managed-block install)
├── INSTALLATION.md ← Detailed install + idempotent merge guide
├── LICENSE ← MIT
├── package.json ← Project metadata (configuration bundle, not an npm plugin)
├── opencode.json ← Config snippet to merge into your opencode.json
├── agents/ ← Agent definitions (OpenCode auto-discovery)
│ ├── mythos-singleshot-thinking-intelligence.md (legacy / fallback)
│ ├── mythos-executor.md (legacy)
│ ├── mythos-verifier.md (legacy)
│ ├── mythos-adversary.md (legacy)
│ ├── mythos-synthesizer.md (legacy)
│ ├── reliability-scout.md
│ ├── reliability-spec-critic.md
│ ├── reliability-test-designer.md
│ ├── reliability-lead.md
│ ├── reliability-verifier.md
│ └── reliability-adversary.md
├── core/ ← Schemas + runtime rules
│ ├── runtime-rules.md
│ ├── routing.md
│ ├── evidence-ledger.md
│ ├── task-contract.schema.json
│ └── verification-report.schema.json
├── skills/
│ └── fable-mythos-modus/
│ └── SKILL.md ← Behavioral priming skill
├── docs/
│ ├── MYTHOS-SYSTEM-CARD-ANALYSIS.md ← Evidence base
│ ├── ANTI-CONCEALMENT.md ← Why every uncertainty is surfaced
│ ├── FAQ.md ← Common questions
│ ├── RELIABILITY-ROADMAP.md ← P2/P3 goals
│ └── EMPIRICAL-BENCHMARK-PLAN.md ← Validation plan
└── diagrams/
└── map-pipeline.svg ← High-res pipeline diagram
| Claim | Confidence | Basis |
|---|---|---|
| The harness enforces a deterministic Done Gate | High | Defined in core/runtime-rules.md; machine-observed |
| Orthogonal scouts produce more diversity than 3 identical clones | Medium | Established ensemble principle; not yet measured here |
| Self-testing + clean-checkout verification is structurally stronger than self-exemption | Medium | Aligns with published multi-agent findings |
| Empirical improvement over a GLM-5.2 baseline | Unproven | Benchmark planned, not yet run — see docs/EMPIRICAL-BENCHMARK-PLAN.md |
Honest limits: This is Mythos-inspired, not Mythos-identical. Prompts can steer behavior; they do not move weights, post-training, or latent representations. Same model = shared blind spots. The harness reduces, does not eliminate, correlated errors. Empirical validation is planned, not yet measured. Anyone claiming "100% Mythos", "1:1", "MAP-v2", "Cybench 100%", "−50–80% hallucination rate", or "★★★★★ output reliability" on this package is misrepresenting it.
This package runs alongside Oh My OpenAgent without conflicts. The agents are independent of OmO's Sisyphus/Oracle/Prometheus fleet. You can use both simultaneously — OmO for its specialized agents, Fable & Mythos for the reliability harness.
Is this affiliated with OpenCode, xAI, ZAI, or any AI lab?
No. This is an independent project. "Mythos" is used as a reasoning-pattern label (not a product claim). OpenCode CLI is an open-source tool. This package is a third-party configuration bundle.
Does this work with Oh My OpenAgent?
Yes. The agents are independent and run alongside OmO's agent fleet. No conflicts.
Does this make OpenCode identical to another model?
No, and we don't claim it will. Observable behavioral patterns transfer. Latent internal processes (SAE features, evaluation-awareness vectors, emotion/persona vectors) are architecture-specific to other models' weights and do not transfer. Net result: a structurally stronger reliability harness, not model parity.
Where are the empirical numbers?
Planned, not yet measured. See docs/EMPIRICAL-BENCHMARK-PLAN.md for the four-variant comparison plan (GLM-5.2 baseline / current harness / compact prompt / harness v2) with the metrics that matter (false_done_rate is the most important).
- fable-mythos-zcode — Companion package for ZCode (GLM-5.2 / ZAI)
- fable-mythos-grok — Companion package for Grok Build CLI (xAI)
MIT — use it, fork it, build on it.
Star · Fork · Install · Roadmap · Benchmark plan
Built on the principle that reliability comes from evidence, orthogonal verification, and machine-observed gates — not from pretending a prompt moves weights.