Model Experience

What the plugin's commands put into model requests, in the DeepSeek Harness package-README format. The package README links here; this document is the full contract.

/compact-directive summarization call

What the model sees

One user message with no system prompt, no tools, and no conversation prefix: the requirement (when given) layered over a four-point baseline (task goal / key steps done / key findings / next step), followed by the middle span rendered as role-prefixed plain text — [user] …, [assistant] …, [tool] called tool …, [tool] tool result:, <image> for images. The baseline requires every [user] line verbatim, and the rendering preserves it.

Token effect

Data-dependent: input ≈ the rendered middle span (bounded by the planner's keep-head/keep-tail defaults of 3 user turns each, and the 4M-character render guard); output ≤ the configured maxTokens.

KV Cache effect

An independent model request: a single fresh message with no conversation prefix, so it reuses no warm-prefix KV cache — the deliberate trade of cache reuse for a more focused summary, mirroring the upstream ContextForge calls. The call invalidates nothing in the session prefix.

/trim-directive chunk calls (operation mode)

What the model sees

One user message per chunk, in parallel: the trim instruction (the requirement is the sole instruction — no four-point baseline) followed by the chunk rendered as numbered nodes — [seq <global event seq>] [role] …, continuation lines indented, reusing the harness's global event-seq vocabulary (session-query references events the same way). The model replies with an operation manifest (delete: / delete-text: / rewrite: / summarize: lines, ---content--- … ---end--- content blocks for rewrite/summarize, or exactly <<NO_CHANGE>> for an unchanged chunk) or the chunk rewritten in full; the command layer executes either (see the README's Operation mode).

Token effect

Input per chunk ≤ 50K heuristic tokens (max 20 chunks; total input beyond 20 × 50K fails loud before any call opens). Output ≤ min(window/2, 256K) per chunk; a chunk that declares <<NO_CHANGE>> costs ~0 output tokens and its original rendering is kept verbatim. A trim that touches a few nodes of a large session pays only for the changed chunks.

KV Cache effect

Independent model requests: every chunk is a single fresh message with no conversation prefix, so parallel chunk calls reuse no warm-prefix KV cache and invalidate nothing in the session prefix.

The checkpoint (what replaces the trimmed span)

What the model sees

The checkpoint lands on the session surface as one user message: a marker naming the requirement ([Directive-driven compaction checkpoint, per requirement: <requirement>], trim variant for /trim-directive), a guard stating that removed content is final and must not be reconstructed, and the summary text with [part N/M] dividers when chunked. On its next request, the model sees the checkpoint in place of the replaced nodes; the UI transcript keeps the original dialogue by dsh's design.

Token effect

Replacement: the shadowed span's tokens give way to the checkpoint's tokens; the shrink validation requires the checkpoint to be smaller than the span it replaces.

KV Cache effect

Replacement of earlier request tokens: the checkpoint becomes part of the session prefix for subsequent requests, which then resume append-only growth as the conversation continues. The summarization and trim calls themselves are independent requests (see above).