dsh-think-strip

dsh-think-strip

Keep DeepSeek Harness conversations alive for hundreds of turns without paying for the same reasoning text twice.

This plugin strips historical reasoning blocks and oversized tool outputs out of the model's context. Originals are archived to disk by timestamp with a searchable index; in their place the model sees a one-line summary (for reasoning) or a head/tail preview plus the archive path (for tool results).

Two per-session switches sit left of the chat composer — thinking and tools — and they are independent. New sessions follow configured defaults.

Interface language: the plugin ships English and Chinese strings (locale/en.json, locale/zh.json). Documentation: English · 中文

---

The problem

Every request DeepSeek Harness sends includes the full history: assistant reasoning blocks are replayed verbatim (no filter in dsh-llm-deepseek's assistant()), and tool results are resent whole. Nothing is wrong with that per se — it is what makes an agent coherent. It is also what makes long sessions expensive and eventually impossible.

Measured on one real session (session-f0b74f94, 1101 tool results / 1118 assistant messages):

A handful of large tool outputs dominate the budget. Strip those 128 results and you reclaim ~1.22M characters; strip the reasoning and you reclaim another ~1.08M. Neither number is a rounding error.

What it guarantees

• The append-only session log is never rewritten. Originals stay in session.v4.jsonl.zstd, so fork / resume / replay behave exactly as before.
• Prompt-cache prefixes survive. Stripping happens mid-history, but once a strip is committed the message sequence is byte-stable, so prefix caching keeps hitting.
• Nothing is lost. Each stripped item is written as its own JSON file, indexed in index.jsonl, and the replacement text in context names the archive path — the model can go read it when it actually needs the detail.

How it works

Two cooperating channels, neither of which touches the log:

1. appendPluginRecord(session, 'plugin:think-strip', { targets, indexed }) — writes one ignorable: true plugin record. It never enters the surface nodes and survives a DSH restart.
2. ctx.sessions.registerMessageProjection({ type: 'plugin:think-strip', project }) — registers the matching projection. On fold, the seqs named by the record are replaced with frozen copies (each keeps its original message.id).

Each entry in targets is { seq, kind, summary?, replacement? }, where kind is 'reasoning' or 'tool'. Records without a kind are read as 'reasoning', so records written by v0.1 keep working and old sessions do not break on upgrade.

Timing: on agent/turn-stopping (the turn is closing and the model owes no reply) the plugin rescans all history, archives originals first, then commits a single record covering every target seq.

Rescanning everything rather than only the current turn matters: turns from before the plugin was enabled — or before a switch was flipped — have accumulated originals that would otherwise sit in context forever. Already-processed messages come back from session.deriveEventMessage(event) without their originals (the projection is applied), so they are skipped for free. session.surface.nodes doubles as a whitelist, skipping messages already hidden by compaction.

Why not some other seam

The two session switches

Two role="switch" buttons left of the composer:

• Thinking — whether this session strips historical reasoning content
• Tools — whether this session compresses oversized historical tool output

State is stored per sessionId in <archiveDir>/switches.json:

{
  "defaults": { "reasoning": true, "tools": false },
  "sessions": { "session-f0b74f94": { "reasoning": false } }
}

Precedence: session override > defaults > Config. A switch only decides whether this turn commits new strip records; records already committed keep applying forever (the log cannot be rewritten — that is the design). To undo stripping for a session you must archive that session, or accept that it stays projected.

Query and update over HTTP (loopback only):

# read
Invoke-RestMethod 'http://127.0.0.1:<DSH_WEB_PORT>/think-strip/state?sessionId=session-f0b74f94'

# turn off reasoning stripping for this session;
# tools:null clears the override and falls back to the default
Invoke-RestMethod 'http://127.0.0.1:<DSH_WEB_PORT>/think-strip/state?sessionId=session-f0b74f94' `
  -Method Post -ContentType 'application/json' -Body '{"reasoning":false}'

Archive layout

<archiveDir>/
├── switches.json                    # per-session switches
├── index.jsonl                      # one index line per archived item
├── plugin.log                       # startup marker (did apply even run?)
└── <sessionId>/
    ├── 2026-10-20T10-33-24-123Z-t3-s1-q42.json      # reasoning
    └── 2026-10-20T10-33-24-456Z-tool-97.json        # tool output

One archived reasoning item:

{
  "kind": "reasoning",
  "time": 1760950000000,
  "sessionId": "session-f0b74f94",
  "turn": 3,
  "step": 1,
  "seq": 42,
  "model": "deepseek-v4.1-flash",
  "chars": 2514,
  "summary": "User wants 1+1.",
  "reasoning": "(full original reasoning)",
  "text": "(the reply text for that message)"
}

One archived tool result:

{
  "kind": "tool",
  "time": 1760950000000,
  "sessionId": "session-f0b74f94",
  "turn": 3,
  "step": 1,
  "seq": 97,
  "toolName": "pwsh",
  "chars": 49932,
  "summary": "pwsh output, 49932 chars",
  "text": "(full tool output)"
}

An index line:

{"time":1760950000000,"sessionId":"session-f0b74f94","kind":"tool","turn":3,"step":1,"seq":97,"toolName":"pwsh","chars":49932,"summary":"pwsh output, 49932 chars","file":"session-f0b74f94/2026-10-20T10-33-24-456Z-tool-97.json"}

Searching it:

# by keyword
Select-String -Path "$env:USERPROFILE\.dsh\think-archive\index.jsonl" -Pattern "some keyword"

# tool outputs only, largest first
Get-Content "$env:USERPROFILE\.dsh\think-archive\index.jsonl" |
  ForEach-Object { $_ | ConvertFrom-Json } |
  Where-Object { $_.kind -eq 'tool' } |
  Sort-Object chars -Descending |
  Select-Object -First 20

What the model sees in place of an oversized tool result:

(head N characters)
…[tool output archived]…
(full 49932 characters archived at: <DSH_HOME>\think-archive\session-xxx\2026-….json; read or grep it)
…(tail M characters)

It reads that file when it needs the detail. What is saved is the indiscriminate context cost, not the information.

Install

From inside DSH:

plugin_manager action=install_bundle target=<absolute path to this package>

Or add it to the profile's cordis.patch.yml by hand:

- insert:
    - id: dsh-think-strip
      name: 'dsh-think-strip'

The client half is declared by package.json's dsh.client.platform = "web" plus exports["./client"]; the host injects it into the Web UI automatically.

Configuration

Set it in the profile's cordis.patch.yml, overriding by id:

- id: dsh-think-strip
  config:
    enabled: true
    archiveDir: ''
    archive: true
    summaryMode: first-line
    summaryMaxChars: 100
    minReasoningChars: 0
    keepSummary: true
    reasoningDefault: true
    toolsDefault: true
    toolStripChars: 4000
    toolHeadChars: 1200
    toolTailChars: 400
    trace: true

Caveats

• The projection lives in the plugin. Uninstalling it makes sessions that used it refuse to derive further (session message projection "plugin:think-strip" was removed or replaced). To uninstall, archive those sessions first or accept that they can no longer be opened.
• One record per turn, not per step. Multiple steps in a turn are covered by the same record.
• Tool results are only processed when they are entirely text blocks. Results containing images / files / structured blocks (32 of 1101 measured were text,image) are skipped whole — splitting image blocks risks more than it saves.
• Content that fails to archive is not projected. Those seqs stay in context with a warning logged — better to spend the tokens than to lose the original.
• When compaction replaces a span of history with a summary, the hidden seqs are no longer on the surface and the projection silently skips them.
• Plugin records must contain no undefined values. Session.append validates with snapshotJsonValue, which only accepts booleans / strings / finite numbers / plain objects / plain arrays; a single undefined throws session event "…" carries non-JSON-serializable data and the projection never commits (the archive is still written, but context is unchanged). Everything therefore goes through toJsonSafe() (lib/pure.js), which strips undefined keys via a JSON round-trip.
• Changing plugin source requires a DSH restart. Modules under node_modules do not participate in HMR; only config changes hot-apply.
• Lifecycle markers: <archiveDir>/plugin.log gets one line per apply; with trace on, each turn appends turn=N sid=… sw=r1t1 ev=<events> sf=<surface nodes> targets=<hits> committed|targets=0|FAILED, which separates "plugin never started" from "started but never fired" from "fired but matched nothing" from "matched but failed to commit".

Tests

The pure logic layer (lib/pure.js) runs without Harness:

cd dsh-think-strip
node --test test/pure.test.mjs

License

MIT — see LICENSE.