# dsh-resume-turn

DSH plugin: **resume an interrupted reply from its partial output** instead of restarting from scratch.

[简体中文](./README.zh.md) · [Repository](https://github.com/Harris-Logic/dsh-resume-turn) · [Issues](https://github.com/Harris-Logic/dsh-resume-turn/issues) · MIT

When a model request fails mid-stream on a slow / flaky endpoint (e.g. the AMD
DeepSeek endpoint — slow inference, frequent disconnects, `TIMEOUT` /
`TRANSPORT` / `SERVER` / `RATE_LIMIT`), the default `dsh-llm-retry` rebuilds
the *same* request and re-runs it **from zero**: everything that had already
streamed (including a long thinking chain) is thrown away, and a multi-minute
attempt can be repeated several times before giving up.

`dsh-resume-turn` changes that: at the failure point it collects the partial
output that **already streamed** (visible text + reasoning chunks) from the
session event log, injects a visible "auto-resumed" steering message that
quotes that partial output and instructs the model to continue from where it
stopped, and takes over recovery with `{ kind: 'retry' }` — so the next
request starts from the interruption point, not from zero.

## How it works

```
agent/request-error (waterfall)
  ├─ user cancelled?        → delegate (never auto-resume over a user stop)
  ├─ code not transient?    → delegate
  ├─ no partial output?     → delegate (nothing to resume; default retry is right)
  ├─ mid tool-call?         → delegate (restart is safer for tool arguments)
  ├─ resume budget spent (default 3 / turn)? → delegate
  └─ otherwise
       ├─ collect current-attempt chunks (assistant/chunk events after the
       │  last attempt boundary) → partial text + reasoning
       ├─ cancellable backoff (2s, doubling, cap 30s)
       ├─ agent.steer(resume message)   # visible "auto-resumed" row; carries
       │                                 # the partial output + "continue" instruction
       └─ return { kind: 'retry' }      # retry rebuilds the request from the
                                         # surface, which now includes the resume message
```

## Install (one command)

```bash
# from GitHub
npx @deepseek-ai/dsh plugin --profile web add github:Harris-Logic/dsh-resume-turn

# or from a local checkout
npx @deepseek-ai/dsh plugin --profile web add /absolute/path/to/dsh-resume-turn
```

Restart the profile (or the web host) to activate. The package declares
`dsh.bundle.patch`, so the installer appends it to `dsh.profile.bundles` and its
`cordis.patch.yml` plugin row is applied automatically — **no manual file
edits**. `@deepseek-ai/*` modules come from the host (the very `schemastery` /
`dsh-llm` instances DSH itself runs on): they are declared as optional peers and
are never installed separately.

## Config

The plugin has no required config. Optional keys (in the profile's
`cordis.patch.yml` row `config`, or as bundle config):

| key | default | meaning |
|---|---|---|
| `maxResumesPerTurn` | `3` | max auto-resumes per turn; beyond that, default retry takes over |
| `resumeCodes` | `["TIMEOUT","TRANSPORT","SERVER","RATE_LIMIT"]` | failure codes eligible for resume |
| `initialDelayMs` | `2000` | backoff before the first resume (doubles per attempt, cap 30s) |
| `maxPartialChars` | `12000` | max characters of partial output quoted into the resume message |

## Interaction with dsh-llm-retry

- On providers whose `retryPolicy` is `mode: always` (recommended for flaky
  endpoints; see the [A1 config](./README.zh.md) notes), `llm-retry` consults
  downstream recovery first — this plugin gets the first refusal. Returning
  `{ kind: 'retry' }` short-circuits the chain; returning `undefined` lets
  `llm-retry` schedule its own backoff.
- On `mode: normal` providers this plugin is only consulted if it is ordered
  before `llm-retry` in the waterfall; otherwise the default restart-based
  retry applies unchanged.

## Limitations

- Resume quotes the partial output into context; the continued answer may
  rephrase a boundary sentence (mitigated by the "don't repeat" instruction).
- If the endpoint fails again right after a resume, attempts count against the
  per-turn budget; after the budget is spent the default retry takes over.
- A host restart resets in-memory resume budgets/boundaries (no custom session
  events are appended, so stored logs stay fully readable).

## License

MIT