| name | harness |
| description | Maintain Gradient's repo-local agent harness: skills, agents, settings,
repo brief, and gate-aligned instructions. Trigger: /harness, /skill,
/primitive.
|
| argument-hint | [create|eval|lint|sync|audit] [target] |
/harness
Harness work in this repo is itself a Gradient Harness artifact. Keep it
cross-harness, public-safe, and honest about the absence of app CI.
Layout
- Shared skill root:
.agents/skills/
- Shared agent root:
.agents/agents/
- Claude bridge:
.claude/skills/
- Codex bridge:
.codex/skills/
- Pi bridge:
.pi/skills/
- Claude agent bridge:
.claude/agents/
- Native manifest:
harness/default/manifest.json
- Machine brief:
docs/gradient-machine-brief.md
Rules
- Shared skills are canonical; harness-specific dirs are symlink bridges.
- Shared agents are canonical; harness-specific agent dirs are projections.
- Substantial workflow skills must declare triad handling with a
## Triad Use
section or an explicit skip rationale. Use references/triad-loop.md for the
planner, executor, critic, operator decision, and evidence contract.
- Workflow skills must cite this exact gate: The load-bearing ship gate is
./scripts/validate.sh, plus reviewer judgment for lifecycle semantics in
docs/architecture.md and docs/module-contracts.md when docs change.
- Universal skills may remain verbatim.
- Repo-tailoring logic must be source-backed. Do not hardcode product gates,
invariants, or workflow doctrine from a language, framework, or prior incident.
Extract them from target repo guidance, committed executable gates, selected
profiles or policy packs, or operator input. Unknowns become readiness work.
- Supply-chain release-delay guidance must come from the selected profile,
standards manifest, policy pack, target repo guidance, or operator input.
Harnesses may instruct agents to use the configured package-manager adapter,
but must not infer npm, Bun, pip, Cargo, Maven, Go, container registry, or
other ecosystem commands from a package manifest alone.
- Repo-tailoring changes must include contrasting fixtures: one documented
nonstandard gate and one sparse repo proving Gradient does not guess from a
package manifest alone.
- Do not add fake command allowlists to
.codex/config.toml.
- Do not depend on Spellbook paths or
.spellbook files for the native default.
- Preserve public-safe boundaries in examples and skill text.
Audit
Check bridge resolution, native agent presence, resolved harness manifest,
settings syntax, public-safe boundaries, and gate sentence drift.
Triad Use
Use the planner/executor/critic triad when harness work changes skill
contracts, generated harness artifacts, policy/context/work/fleet module
boundaries, repo-tailoring behavior, or user-facing workflow guidance.
Planner must produce:
- goal and non-goals;
- repo anchors and authority order;
- executor ownership boundary;
- executable evidence oracle.
Executor must:
- stay inside the ownership boundary;
- use the repo's normal edit/test workflow;
- report changed files and evidence.
Critic must:
- review cold against the planner packet;
- fail on missing evidence, module-boundary drift, unnecessary complexity, or
unapproved scope growth;
- give concrete fixes for every blocker.
The lead agent/operator decides whether to accept, loop, or stop.
References
references/triad-loop.md: provider-neutral planner/executor/critic loop
for substantial skills.