Files
deepseek-harness/packages/spill/spill-policy
Tianyi Cui 6520f71f94 Merge remote-tracking branch 'origin/codex/invariant-service-seam' into codex/invariant-package-registration-gate
# Conflicts:
#	.agents/notes/implemented/architecture/2026-07-19-package-owned-invariant-service.i18n.yaml
#	docs/core-data-structures/session.md
#	docs/module-graph.md
#	package.json
#	packages/sdk/scripts/tsconfig.json
2026-07-20 00:35:06 +08:00
..
2026-07-19 22:52:03 +08:00

@deepseek-ai/dsh-spill-policy

The tool-result spill policy: a tools/post-execute transformer that keeps oversized plain-text tool results out of the model's context. When a final result exceeds maxInlineBytes, it saves the FULL text through ctx.spillStore and replaces the model-facing result with a bounded head/tail preview plus the backend's locator and retrieval hint.

This plugin registers no service and owns no storage or preview mechanics: preview is @deepseek-ai/dsh-retention (TextRetainer), storage is ctx.spillStore. It only decides WHEN to spill and composes the notice.

Config

Key Default Meaning
maxInlineBytes (omitted) Model-facing context cap for a plain-text result, in UTF-8 bytes (a non-negative integer; validated at load). Omitted disables the policy entirely (the plugin registers nothing). When set, a larger result is spilled and replaced with a preview derived from the same budget (head/tail split).

Behavior

  1. Let the tool run (delegates via next(), so it bounds whatever a downstream hook accepted).

  2. Skip read (avoids a read → spill → read again loop) and any non-accept decision (a block's corrective feedback passes through).

  3. Flatten the accepted content only when it is plain text (all text blocks); a result with any non-text block is left untouched.

  4. If its UTF-8 size is ≤ maxInlineBytes, leave it unchanged.

  5. Otherwise save the full text and replace the result with a preview + this notice, sized so the whole replacement (preview + blank line + notice) stays within maxInlineBytes — the notice's byte cost is reserved out of the budget, so the preview shrinks to fit and the model-facing result never exceeds the cap:

    <retained head/tail preview>
    
    (Omitted N bytes. Full formatted result stored at: /…/session-…/…-web_fetch.txt. Use read with offset/limit, or grep this path to search within it.)
    

    When the notice alone fills the budget (a tiny cap or a long locator) the preview is empty and only the notice is returned. If even that notice-only replacement would exceed maxInlineBytes, the policy keeps the inline result — it never emits a replacement over the cap (and a within-cap replacement is always smaller than the original, so this also means spilling never adds bytes).

Best-effort: no session owner, no ctx.spillStore backend, or a saveText rejection ⇒ the policy logs a warning and returns the original result. A spill failure never turns a successful call into an isError or hides the inline result.

Scope

The policy sees only the FINAL formatted tool result — not a tool's internal resource. If a provider already truncated (e.g. web-fetch-local.maxBodyChars), the spill artifact holds the full formatted result the tool returned, not the full original source. Provider/resource caps stay mandatory and separate. Tool-owned early spill (bash streams, subagent rollouts) is future work — see the tool output spill Agent Note.

Model Experience

Oversized plain-text result

What the model sees

Results at or below maxInlineBytes, read results, blocked decisions, and results containing non-text blocks are unchanged. An oversized plain-text result becomes a bounded head/tail preview followed by (Omitted <bytes> bytes. Full formatted result stored at: <locator>. <retrievalHint>); storage or ownership failures leave the original result visible.

Token effect

A successful replacement is at most maxInlineBytes UTF-8 bytes and remains in history until compaction; the full spill text is not resent to the model.

KV Cache effect

Append-only; newly visible content follows the reusable request prefix and does not invalidate existing KV-cache entries.

Known Limitations and Deferred Work

  • Only final plain-text results are spillable — mixed-content results, blocked feedback, and read pass through; provider truncation or tool-owned retention that happened earlier cannot be recovered here.
  • A notice that cannot fit disables replacement for that call — a tiny cap or long locator leaves the oversized original inline after the backend has already saved an unreferenced spill.