Compaction Router for DeepSeek Harness

Reroute /compact (and session-title) requests to a second, wider-context model, so a large session's summary never gets truncated just because the live chat model's own context window is too small to hold the replayed history plus the summary output.

Why this exists

Compaction replays the full history region into a request sent to the live model, tagged purpose: 'compaction'. On a large session the replayed history alone can consume nearly all of that model's context — leaving only a few thousand tokens for the summary itself, which then gets cut off (finish_reason: "length") no matter how high maxTokens is set in the compaction config. Raising maxTokens does not help: the ceiling is the model's total context, not the output budget. Pointing compaction at a model with a genuinely larger context window is the actual fix.

Install

Published on npm as @phamkhachoabk/dsh-compaction-router. Or build a tarball from a checkout:

pnpm install
pnpm run build
npm pack
dsh plugin --profile web add /abs/path/phamkhachoabk-dsh-compaction-router-0.1.0.tgz

Restart the harness afterwards: bundle layers are composed at boot.

Configure

| Setting | Default | Meaning | | --- | --- | --- | | provider | gpt9router | Provider route for the compaction target model, exactly as configured on Settings → Models. Leave empty ("") to disable rerouting entirely. | | model | cc/claude-sonnet-5 | Model id at that route. Needs enough context to hold the replayed session history — see "Choosing a target model" below. | | maxTokens | 65536 | Output budget for the summary, replacing compaction's own default. |

Choosing a target model

The route and model need real, verified headroom over your largest sessions' replayed-history size — not just "bigger than the default context." Undersized targets fail exactly the same way the live model did. Verify a candidate against your largest session before trusting it in production.

How it intercepts

Same seam as @phamkhachoabk/dsh-vision-bridge: the harness's llm/stream waterfall, the last boundary before a request leaves for the provider. Only requests tagged purpose: 'compaction' are touched — ordinary conversation requests and purpose: 'session-title' requests pass straight through untouched, since a title generation call is small and does not need the wider context. A matching request is re-dispatched once with provider/model/maxTokens overridden; the rerouted copy is marked (module-local WeakSet) so the interceptor recognizes it on the second pass — the one its own re-dispatch causes — instead of rerouting it again.

Composes with vision-bridge regardless of registration order. Both plugins rewrite by fully re-dispatching through ctx.llm.stream(...) rather than mutating the request mid-chain via next(), so each rewrite restarts the whole llm/stream waterfall from the top. A compaction request carrying images gets its images stripped and its target rerouted in either order — each plugin's own guard makes its second pass a no-op.

Cosmetic note: the harness's own compaction/summary log event records the provider/model from the engine's configuration (the live model), not the rerouted target, since the interceptor only changes what actually gets called — not what the surrounding compaction code believes it called. The event's token usage, however, is the real usage from the rerouted call. Expect the provider/model fields in that log line to look wrong; they are cosmetic only.

Development

pnpm install
pnpm run build
pnpm test