dsh-max-token-auto-continue

Readme

# dsh-max-token-auto-continue

English | [日本語](README.ja.md)

A community host plugin for DeepSeek Harness (DSH) that automatically continues a root-agent session after a turn ends because the model reached its maximum output-token limit.

> This is a community-maintained plugin, not part of the DeepSeek Harness core distribution.

## Behavior

When a root-agent turn ends with `turn/end` reason `max-tokens`, the plugin waits for the agent to become idle and then continues through the appropriate DSH path:

- normal session: `agent.followup()`;
- active, disarmed `/goal`: `GoalService.resume()`.

The continuation is bounded by `maxConsecutive` and is cancelled when newer human input or a newer turn supersedes the pending continuation.

## Safety properties

- Root agents only; subagent turns are not auto-continued.
- Human input is not impersonated. Generated follow-ups use `source.kind: "dsh-max-token-auto-continue"` and `form: "notice"`.
- No action is taken for paused, blocked, complete, or active+armed goals.
- Pending continuation is invalidated by newer human input, a newer turn, plugin unload, or plugin disable.
- Failures stop the continuation path instead of retrying indefinitely.
- No persistence, UI automation, background network retry, or external network access is implemented.

## Configuration

| Setting | Default | Meaning |
| --- | ---: | --- |
| `enabled` | `true` | Enables automatic continuation |
| `maxConsecutive` | `3` | Maximum automatic continuations in one consecutive max-token chain |

The continuation text and delay are intentionally fixed rather than configurable.

## Compatibility

Latest tested DSH version: **0.2.1-alpha.1**.

| DSH version | Verification |
| --- | --- |
| 0.1.6-alpha.2 | Supported compatibility target |
| 0.1.7-alpha.2 | Plugin load and inventory visibility verified |
| 0.1.7-rc.1 | Plugin load and inventory visibility verified |
| 0.1.7-rc.2 | Plugin load and inventory visibility verified |
| 0.2.0-rc.1 | Plugin load, inventory visibility, and real max-token auto-continue E2E verified |
| 0.2.0-rc.2 | Plugin load and inventory visibility verified |
| 0.2.1-alpha.1 | Build, 13 tests, Web-profile config output, and startup load verified |

Actual auto-continue and Goal resume firing on **0.2.1-alpha.1** has not yet been re-verified. Newer DSH versions are not assumed compatible until verified.

## Install

Install dependencies and build:

```sh
npm install
npm run build
npm test
```

Add the local checkout to the target profile:

```sh
dsh plugin --profile <profile> add <absolute-path-to-this-repository>
dsh --profile <profile> --dump-config
```

For the Web profile, replace `<profile>` with `web`.

The bundle declaration lets DSH add the required profile entries. To override plugin settings, add a config entry with id `dsh-max-token-auto-continue` to the target profile's `cordis.patch.yml`.

## Operational note

This plugin intentionally continues work after an output truncation without waiting for another human message. That is useful for long agent tasks, but it can also extend an unintended task if the model reaches `max-tokens` while already heading in the wrong direction.

The main bounds are root-agent-only scope, supersession by newer input/turns, fail-closed behavior, and `maxConsecutive`.

## Build and test

```sh
npm run build
npm test
```

TypeScript is compiled to `lib/`; the distributable plugin entry is `lib/src/index.js`.

`@deepseek-ai/dsh-llm` is resolved at runtime from the DSH process entry so the plugin uses the host runtime copy.

## Privacy

The plugin does not persist session data and does not send data to external services. Its only session mutation is the bounded continuation message or Goal resume call described above.

## Contributing

Bug reports and focused pull requests are welcome. For compatibility reports, include the DSH version, profile, whether the session was a normal session or `/goal`, the observed `turn/end` reason, and whether a continuation was expected.

## License

MIT. See [LICENSE](LICENSE).