# Triton Agent State And Artifact Templates

Keep Triton loop records under `.humanize/triton-agent/` in the target vLLM or
SGLang checkout, in the nearest project root for direct-file mode, or in the
user-provided workdir when the Triton file is standalone. Do not use
`.humanize/lightop-agent/` for Triton work.

## refined-plan.md

```markdown
# Refined Plan

## Target
- Framework:
- Repo path:
- Target mode:
- Target file:
- Project root/workdir:
- Commit:
- Operator/kernel:
- User goal:

## K/R/W/E/B/P
- K:
- R:
- W:
- E:
- B:
- P:

## Baseline Commands
- install:
- import smoke:
- backend proof:
- correctness:
- benchmark:
- profile:

## Candidate Plan
1. Baseline and backend proof.
2. First edit:
3. Validation:
4. Next gate:

## Stop Conditions
- Correctness:
- Performance:
- Profiling:
- User constraints:
```

## backend-map.md

```markdown
# Backend Map

## Framework Or Call-Site Routing
- Entry point:
- Backend registry:
- Env vars:
- CLI args:
- Direct caller/harness:
- Selected backend:
- Fallback backend:

## Triton Kernel
- Python file/function:
- JIT/autotune decorators:
- Config source:
- Grid function:
- Kernel name in profiler:
- Triton cache or dump path:

## Proof
- Command:
- Evidence:
- Caveats:
```

## attempt-ledger.md

```markdown
# Attempt Ledger

| Version | Hypothesis | Files | Correct | p50 us | p90 us | Delta | Profile | Decision |
| --- | --- | --- | --- | ---: | ---: | ---: | --- | --- |
| v000 | baseline | none | pass | | | | | baseline |
```

## kernel_opt_readme.md

```markdown
# Kernel Optimization Report

## Summary
- Best version:
- Baseline:
- Current best:
- Target:
- Selected card:
- Backend:

## Validation Matrix
| Test | Command | Result |
| --- | --- | --- |

## Performance Matrix
| Shape | Dtype | Backend | Config | Baseline us | Candidate us | Delta | Notes |
| --- | --- | --- | --- | ---: | ---: | ---: | --- |

## Profile Evidence
- Baseline artifact:
- Candidate artifact:
- Bottleneck:
- Next edit:

## Rejected Lineages
| Version | Reason |
| --- | --- |
```

## lineage.jsonl

Write one JSON object per benchmarked candidate:

```json
{"version":"v001","parent":"v000","hypothesis":"tune BLOCK_N for decode","correct":true,"decision":"keep","artifact":".humanize/triton-agent/profile-artifacts/v001/digest.md"}
```