dsh-force-compact

Aggressive, local-first context compaction for DeepSeek Harness agents.

A DSH Cordis function plugin that keeps the agent's working context lean by design, so you can deliver a large-window experience against a self-hosted llama.cpp serving Qwen3.8‑27B at modest context — no API cost, no data egress.

中文

---

Why

Most harness setups bolt a big frontier model onto a short context budget. This plugin makes the opposite bet: you own the weights, the endpoint, and the context budget.

• Self-hosted inference. Point the agent at a local OpenAI-compatible llama.cpp server running Qwen3.8‑27B (GGUF / NVFP4 / MTP variants all work through the standard DeepSeek adapter path — no separate llama.cpp adapter needed).
• Low context, high signal. Rather than fight a small hard cap, the plugin shrinks the conversation itself, so the agent reasons over a tight, high-signal prompt while effectively reaching a much larger working memory.
• Thinking-off by default. disableThinking: true turns off reasoning effort on every outbound call, enforced at two complementary seams (real DeepSeek honors one; llama.cpp honors the other — see "Backend-agnostic thinking control" below).
• Private & free. No per-token billing, no egress, and the exact model/context tradeoff is yours to dial.

---

What it does

Two compaction engines coexist behind one facade (resolveCompaction), transparent to callers:

Engine	Used when	Notes
Official	compaction service is resolvable in the agent realm	Preferred; delegates to compaction/basic.
Builtin	Automatic fallback when the service is realm-isolated (typical standard preset)	Self-contained persistent transaction using only ctx.sessions / ctx.llm.stream / ctx.tokenMeter. Reuses the official compaction/* event vocabulary, so it survives cross-build replay with no ignorable hacks.

You never toggle — official wins when reachable, builtin takes over otherwise.

Trigger points

• Per-request guard (agent/pre-step) — reads the session's projected context tokens (the exact number the harness renders bottom-right, provider-anchored). At autoThresholdTokens it rejects the outgoing request and compacts the head instead, retaining the latest retainLatestTokens verbatim.
• Turn-end / idle (agent/status → idle) — when the agent quiesces, optionally compacts via compactNow (gate: turnEndForceCompactionEnabled).
• Manual /force-compact — immediate compactNow when idle; queues a process-local force flag consumed at the next model step when busy.
• session/flush — the awaited durability checkpoint.

Every path funnels into the single "compaction result landed in the session" boundary — the same point where the live-UI signal fires.

Provider-anchored decisions

Decisions key off projectedTokens (same figure shown in the UI corner), so the plugin never drifts from what you see. Threshold-aware shrink gates skip summarizer calls that provably could not pull the session below the threshold (kills the low-threshold dead loop).

Meter-aligned shadow-price billing

The builtin transaction bills shadowedTokenCount from the same tokenMeter.measure per-node prices the official engine uses, so the meter's collapse protocol settles the drop correctly — the bottom-right counter goes down after compaction instead of drifting upward.

Backend-agnostic thinking control

disableThinking is enforced at two complementary seams:

1. Request seam — reasoningEffort:'off' → the DeepSeek adapter serializes thinking:{type:'disabled'} (real DeepSeek APIs honor it).
2. Wire seam (llm/stream) — the plugin appends top-level reasoning_effort:"none" post-serialization, which llama.cpp's OpenAI-compatible layer parses natively (server-common.cpp maps it to enable_thinking=false, independent of template capability). Real DeepSeek endpoints simply ignore the unknown key.

Result: thinking is genuinely off on any backend, with no target-sniffing heuristic to miss a route.

Live-UI status

A tiny host→client messenger (a liveUi settings field mirrored live to the browser) paints a badge beside the turn:

• 🟥 compressing — pinned red [强制压缩中>>>], fired just before a compaction commits;
• 🟢 done — pinned green [压缩完成!], fired the instant a compaction result lands; 3 s later a fresh random "working" pair takes over;
• 🔵 working — otherwise a playful random one-liner ("正在缝合上下文…", "正在憋大招…").

Publishers are fail-safe: a messenger glitch can never disturb the actual compaction.

---

How it works

agent/request(payload, next)              # every model request
    disableThinking? -> { ...config, reasoningEffort: "off" }

agent/pre-step(payload, next)             # before each model step
    projectedTokens >= autoThresholdTokens?
        no  -> next()                     # let the model request proceed
        yes -> compactRegion(head-before-retainLatestTokens, signal)
              return { kind: "reject" }   # no model request this step

agent/status({ agent, status })          # lifecycle transition
    status === "idle" && turnEndForceCompactionEnabled?
        -> compactNow(agent, freshSignal) # turn-end compaction

session/flush(session)                   # durability checkpoint
    select region -> project messages -> preview + shrink gate
    -> compaction.compactRegion(start, end, agent, signal)

Supporting modules:

• src/hooks/guard.js — per-request guard: thinking-off + threshold gate + forced flag.
• src/hooks/command.js — the /force-compact command.
• src/hooks/idle.js — turn-end forced compaction.
• src/hooks/wire-rewrite.js — the llm/stream wire patch appending reasoning_effort:"none".
• src/engine/region.js — head/tail-anchored region selection (with the official pairing ledger).
• src/engine/summarizer.js — the one-shot LLM summarizer (fully aligned with official compaction-basic: target resolution, prefix-cache alignment, purpose:'compaction' tag, fail-closed finish classification, usage capture).
• src/engine/builtin.js — the builtin persistent transaction (official compaction/* vocab).
• src/engine/checkpoint.js — preview + shrink gate + delegation to the compaction service.
• src/core/projected.js — provider-anchored projectedTokens reading.
• src/core/ui-signal.js — the live-UI messenger.

---

Install

As an installable bundle (recommended):

# from npm (published):
npm install @falling-ts/dsh-force-compact
# from git:
dsh plugin --profile web add github:falling-ts/dsh-force-compact
# from a local checkout:
dsh plugin --profile web add ./dsh-force-compact

Or, from a local checkout, as a --patch overlay without installing:

dsh web --patch dsh-force-compact/cordis.patch.yml

Plugin loaded ⟺ ~/.dsh/logs/dsh-force-compact.log gains:

[force-compact] debug logging enabled — writing [force-compact] lines to <absolute path>

Verify a compaction happened:

idle compaction (builtin) shadowed N nodes (~M tokens)
builtin compaction OK — replaced span seq[A..B] (N nodes, ~K tokens) with a P-char checkpoint

---

Settings

$DSH_HOME/settings.yaml, namespace falling-ts-force-compact:

key	type	default	meaning
disableThinking	boolean	true	Disable reasoning effort on every outbound call (both seams above).
autoThresholdTokens	number ≥ 32000	32000	Projected-token trigger for the per-request gate. Lower ⇒ more aggressive. Floor 32000 (stored values clamp back up at read time).
retainLatestTokens	positive int ≥ 8000	8000	Retain the latest N tokens verbatim; send everything older to the summarizer in one batch. Floor 8000. Drives both the auto gate and the /force-compact path.
turnEndForceCompactionEnabled	boolean	true	Compact on the agent's idle transition.
debug	boolean	true	Emit [force-compact] diagnostics to the plugin log.
logFile	string	~/.dsh/logs/dsh-force-compact.log	Diagnostics destination (~ expands to home dir).
compactionMode	'realm' | 'global'	'realm'	Official-service resolution strategy (priority-1 path).
builtinEnabled	boolean	true	Gate for the builtin engine fallback.
maxSummaryTokens	integer (1024–200000)	1024	Cap on the summarizer LLM maxTokens.

Example — an aggressive local profile:

falling-ts-force-compact:
  disableThinking: true
  autoThresholdTokens: 40000   # compact sooner ⇒ keep the live prompt small
  retainLatestTokens: 8000
  turnEndForceCompactionEnabled: true

When the settings service is absent, the plugin falls back to the same defaults and still compacts — the namespace is optional, never a hard dependency.

---

Screenshots

Settings panel — 设置 > 强制压缩. All nine fields above can be edited live without restart.

Conversation page — the live-UI signal paints three states (🟥 compressing / 🟢 done / 🔵 working); the green banner fades after ~3 s back to a random "working" line.

Tuning for low-context llama.cpp

Serve Qwen3.8‑27B with a comfortable-but-modest context and let the plugin decide the effective window: keep autoThresholdTokens comfortably below the served context so the live prompt stays small and latency flat, while the agent retains deep memory through the compressed head. Pressure is measured in projected tokens (provider-anchored), so the threshold maps predictably onto what the UI shows you.

---

Behavior notes

• Runtime dependency: the compaction service (preset plane agent-presets:compaction-basic). Read live via ctx.get('compaction'); when unreachable the forced-compaction path falls through and lets the request proceed.
• Optional dependencies: settings / tokenMeter / commands / llm / agents are read via ctx.get(...) with guards — a missing one degrades gracefully rather than crashing.
• Per-request settings read: parameters are read per model request, so edits take effect on the next request without a restart.
• Signals: the agent/* Waterfalls forward the current turn's signal; the session/flush checkpoint and the agent/status idle listener each mint a fresh AbortController.
• Persistence: the durable output is the compaction/* bracket events + a surfaceOp:replace user/message checkpoint, replay-safe across builds.
• Client half: web/client.js adds a Settings section "强制压缩 / Force Compact" for editing these values live (uSES-safe mirror, no timers/state).
• One intentional timer: the 3 s publishDone fallback (presentation-only, documented deviation). Otherwise the plugin is pure listeners + a process-local Map force flag.

---

License

MIT (see LICENSE).