TT-Wang/dsh-slice-agent-loop

dsh-slice-agent-loop

中文

A conversational-context retention policy for DeepSeek Harness 0.1.7-rc.2 / 0.1.7-rc.1 (session format V4). It runs alongside the stock agent loop, keeping the host's lifecycle, scheduler, inbox, persistence, request-series handling and full request-reconstruction invariant. The patch is additive: it adds this plugin and nothing else.

History model: an append-only tape

At the first step of a turn, eligible assistant/tool spans from completed turns beyond history.keepRecentTurns (default 0) are sealed into frozen [slice tape v1 …] user/message surface replacements at their original positions, with source kind plugin:slice:history. Existing entries are never re-rendered or nested. Sealing preserves the surface prefix through the last existing entry; newly eligible older nodes are left in place rather than backfilled ahead of that prefix. Protected nodes can split a turn into several entries.

Every original human user message still present on the host surface stays verbatim at its original node, in every turn. User text is not copied into the tape. The system prompt (surface node 0, replaced in place only by the host) and every other system/message / developer/message node are never sealed or cited. Instruction messages, multimodal input and the newest runtime-context snapshot (source kind runtime-context) also keep their original sources and positions. Superseded snapshots can be absorbed with their turn only after the frozen prefix. A snapshot that becomes superseded later, before an existing entry, stays raw at its original position. Snapshot text is never rendered as a human request. Raw events remain in the session log and can be recalled after sealing, folding or resume.

Frozen entries preserve their bytes, but this does not guarantee that each request bills only one new entry: a seal can make everything after its position miss the provider cache, including retained raw turns. See Prefix behaviour.

Entry contents

One entry renders, in order:

• a header naming the span: [slice tape v1 · turns N-M · K turn(s) sealed · recall_turn({"turn":"<n>","view":"dialogue"}) returns a turn's dialogue; expand_result({"seq":<q>,"formatVersion":4}) returns a tool result];
• per sealed turn: [turn N], then every visible assistant text message in that span, in source order, each with a reply wrapper and its source locator; by default the entire text is preserved, including whitespace and messages preceding a tool call;
• the turn's read index line (below);
• the turn's tool lines — [tool turn N step S seq Q · <name> · <size> chars · expand_result({"seq":Q,"formatVersion":4})], at most 6 per turn, each pointing at the durable log record instead of repeating the text.