README.md

---
pretty_name: Kimi K3 Coding & Debugging Agent Traces
license: cc-by-4.0
language:
  - en
annotations_creators:
  - machine-generated
task_categories:
  - text-generation
size_categories:
  - n<1K
tags:
  - traces
  - code
  - agentic
  - tool-use
  - coding-agent
  - coding-agents
  - agent-traces
  - sft
  - distillation
  - reasoning
  - chain-of-thought
  - cot
  - moonshotai
  - kimi
  - k3
  - kimi-k3
  - openrouter
  - pi-agent
  - pi-coding-agent
---

# Kimi K3 Coding & Debugging Agent Traces

> Generated by **[moonshiner](https://github.com/greghavens/moonshiner)** — an open harness for
> distilling verified, model-attested agentic coding traces.

Real, end-to-end **agentic coding trajectories** produced by
`moonshotai/kimi-k3` driving the **pi** coding-agent runtime over
`openrouter`, at `max` reasoning. Each trajectory solves a concrete
repair or build task in a real repository — reading, editing, and running code
with tools — and is published only after its work **verifiably passes** — acceptance tests green and protected test files unmodified (hash-checked) — with **dual model attestation**. This is a live, growing drop: each row is one whole-session trajectory, appended as it finishes. At release the corpus is screened by **gpt-5.6-sol** (`codex`) and re-exported as cumulative next-step prefixes (one training row per assistant turn), replacing this file wholesale.

<!-- screened-snapshot:start -->
**Generation in progress:** 106 verified, model-attested trajectories published so far, drawn from a corpus of 883 verifiable repair seeds; this file grows continuously as new traces finish. Model-attested: **100%**. At release the screened next-step export replaces this file and this card regenerates from the published rows.
<!-- screened-snapshot:end -->

## What makes it different

- **Every trajectory verifiably solved its task.** Acceptance tests pass and the
  protected test/spec files are unmodified (hash-checked) before a trace is ever
  eligible — no self-reported success.
- **Independent judge at release.** A separate reviewer (`gpt-5.6-sol`) screens every trajectory before the final export; the teacher does not grade itself.
- **Attested provenance.** 100% of rows are model-attested: both the
  agent event stream *and* a host-side loopback proxy observed the upstream
  answering as the declared model — a sandbox cannot self-certify its identity.
- **Runtime-normalized.** Native runtime traces are converted to a single
  OpenAI-style `messages` + `tools` schema, so the rows are agent-agnostic.
- **One tool surface for every task.** Each trajectory ran with the same offered
  action space (listed below); the release export carries the full JSON tool
  schemas per row.
- **Privacy-scrubbed.** Host paths, runtime scratch directories, and credential
  patterns are redacted; reference answers and provider keys are gated out of the
  export and fail the build if present.

## Task program

Accepted trajectories by category (one count per whole trajectory, not per row):

| Category | Trajectories | Share |
| --- | ---: | ---: |
| `build-game` | 28 | 26% |
| `build-lib` | 10 | 9% |
| `full-distill-brownfield-orientation` | 4 | 4% |
| `build-cli` | 4 | 4% |
| `build-report` | 3 | 3% |
| `warnfix-bash` | 3 | 3% |
| `warnfix-c` | 3 | 3% |
| `build-parsing` | 3 | 3% |
| `compilefix-c` | 3 | 3% |
| `compilefix-cpp` | 3 | 3% |
| `warnfix-cpp` | 3 | 3% |
| `build-fs` | 2 | 2% |
| `syntax-bash` | 2 | 2% |
| `debug-memory` | 2 | 2% |
| `syntax-c` | 2 | 2% |
| `syntax-csharp` | 2 | 2% |
| `debug-cli` | 1 | 1% |
| `build-config` | 1 | 1% |
| `refactor-argv` | 1 | 1% |
| `build-text` | 1 | 1% |
| `build-locking` | 1 | 1% |
| `build-concurrency` | 1 | 1% |
| `escape-wordsplit` | 1 | 1% |
| `build-logs` | 1 | 1% |
| `debug-report` | 1 | 1% |
| `build-docs` | 1 | 1% |
| `build-testing` | 1 | 1% |
| `debug-fs` | 1 | 1% |
| `escape-argv` | 1 | 1% |
| `debug-deploy` | 1 | 1% |
| `escape-ansic` | 1 | 1% |
| `debug-pipeline` | 1 | 1% |
| `build-data` | 1 | 1% |
| `debug-numeric` | 1 | 1% |
| `escape-config` | 1 | 1% |
| `build-datastruct` | 1 | 1% |
| `escape-string` | 1 | 1% |
| `build-binfmt` | 1 | 1% |
| `debug-logic` | 1 | 1% |
| `escape-rawstring` | 1 | 1% |
| `syntax-cpp` | 1 | 1% |
| `build-channels` | 1 | 1% |
| `build-events` | 1 | 1% |
| `project-fulfillment` | 1 | 1% |
| `build-generics` | 1 | 1% |

Spanning domains:

| Domain | Trajectories | Share |
| --- | ---: | ---: |
| `coding` | 106 | 100% |


Multi-step depth: assistant turns per trajectory range 4–20 (median 8).

## Languages

`bash`, `c`, `cpp`, `csharp`, `asm`

## Tool surface

Offered to the teacher on every task: `read`, `write`, `edit`, `bash`, `grep`, `find`, `ls`.
Exercised across the corpus: `bash`, `edit`, `read`, `write`.

## Schema

| Column | Type | Description |
| --- | --- | --- |
| `task` | string | Seed/task id the trajectory solved. |
| `lang` | string | Primary programming language. |
| `category` | string | Task kind (build, debug, feature, refactor, …). |
| `teacher_runtime` | string | Coding-agent runtime that produced the trace. |
| `teacher_model` | string | Teacher model id. |
| `provider` | string | Serving provider for the teacher model. |
| `reasoning_effort` | string | Teacher thinking/reasoning level. |
| `model_attested` | bool | Both the agent stream and the upstream proxy confirmed the model. |
| `observed_models` | list[string] | Model id(s) actually observed in the attested stream. |
| `trace_format` | string | Native runtime trace format before normalization. |
| `n_messages` | int | Message count in the trajectory. |
| `messages` | list[object] | The full session (system/user/assistant/tool) in OpenAI chat format. |

At release, rows switch to the cumulative next-step schema used by the sibling
datasets (`source_trajectory_id`, `split`, `assistant_step`, `target_message_index`,
JSON-encoded `tools`, …).

## Layout

A single file, `traces.jsonl`, holds one row per **whole trajectory** — the full
session in `messages`. The file is rewritten and re-uploaded automatically as new
verified trajectories land. At release it is replaced by the validated cumulative
next-step export: one row per assistant turn, exact-prefix continuity and
trajectory-disjoint train/val splits checked before publication.

## Provenance & reproducibility

- **Seeds.** Tasks are real repair fixtures (prompt, verify command, protected
  test files, reference fix) imported with a deterministic precedence — a
  canonical source, falling back to a secondary only where the canonical lacks a
  seed. The reference fix is used **only** to prove solvability and is never
  exported.
- **Teacher.** `moonshotai/kimi-k3` via `pi` over `openrouter`,
  `max` reasoning. The real provider key stays host-side behind a
  loopback proxy; the sandboxed agent receives only a dummy token.
- **Screening.** Deterministic gates (acceptance tests + protected-file hashes)
  are already applied to every published trajectory; the independent `gpt-5.6-sol`
  review gates the final release. Standing rejections may be retraced and
  re-screened.
- **Derivation.** `whole-session trajectories (release: cumulative-next-step-prefixes)`
  — validated for exact cumulative-prefix continuity and trajectory-disjoint
  splits at export time.

## Limitations

- **Success-filtered.** Only solved, screened trajectories are included, so the
  data reflects successful problem-solving, not the full distribution of attempts
  or failure recovery.
- **Judge bias.** Screening reflects the judge model's preferences; a second
  model's blind spots carry through.
- **Pre-screening drop.** Every published row passed deterministic verification
  and attestation, but the independent judge screen happens at release; a small
  share may be removed or replaced in the final export.
- **Reasoning is as-emitted.** Only reasoning the teacher surfaced in-stream is
  present; no hidden chain-of-thought is reconstructed.

## Intended use

- **SFT / distillation** of agentic coding models on verified, multi-step,
  tool-using trajectories.
- Studying **next-step planning** and tool-use behavior in real repositories.
- Behavioral analysis of a strong teacher under deterministic verification.

## License

Released under **cc-by-4.0**. If you use this dataset, please attribute it:

> Kimi K3 Coding & Debugging Agent Traces — https://huggingface.co/datasets/greghavens/kimi-k3-coding-and-debugging-traces