dsh-leet-accelerator

Independent, optional allocation-conscious tool-result pruning for DeepSeek Harness (DSH), with per-optimization switches and Web UI buttons. This project is not related to OpenBot and is not an official DeepSeek plugin.

中文说明 · Verification record

What it changes

Upstream @deepseek-ai/dsh-compaction-tool-result-pruner implements measureContent and pruneContent by expanding each text block with Array.from(text) into a code-point array, then slice().join('') on the head and tail. For a long tool result that allocates one array of per-code-point strings plus two more slices.

This bundle replaces the toolResultPruner provider only where its patch applies and keeps the upstream pruneSession (shadow-price events, token meter, replay). Two independent switches:

Switch	Default	What it does
enabled	false	Linear UTF-16 scan instead of Array.from, recording in the same pass whether the block contains a valid surrogate pair. Boundary lookup is O(1) when there is no pair (a UTF-16 offset equals a code-point position), otherwise it scans from the nearer end. No whole-string code-point array is allocated.
fastLength	false	Experimental candidate. Probes with a regular expression; when the block has no adjacent surrogate pair, text.length is already the code-point count, so the linear scan is skipped. Text with pairs pays for the probe and still runs the scan. Depends on enabled; with enabled: false it does not participate in the algorithm.

Four combinations are tested: enabled × fastLength. In either enabled: false state the code delegates to super.measureContent / super.pruneContent, i.e. the official algorithm runs unchanged and context protection stays on.

fastLength is not a validated optimization. It regresses on emoji-heavy and multi-block inputs (see below) and has never been measured against a real model task. Keep it off unless you have measured your own workload.

Local microbenchmark

benchmark/compare.js, Node v24.16.0 on darwin/arm64, pruneContent only. 12 warmups, 60 alternating samples per variant, alternating variant order each round. Milliseconds. Full data in benchmark/last-run.json.

Case	Code points	Official P50 / P95	Linear P50 / P95	+fastLength P50 / P95	Linear gain	fastLength vs linear
short-ascii	4096	0.0261 / 0.1252	0.0075 / 0.0630	0.0003 / 0.0045	71.2%	95.6%
large-ascii	300000	6.5405 / 16.5320	0.8298 / 3.1730	0.0031 / 0.0075	87.3%	99.6%
large-cjk	300000	6.5791 / 34.3798	0.8538 / 3.3254	0.2352 / 0.8321	87.0%	72.5%
large-emoji	150000	5.6723 / 20.9982	0.7117 / 3.0680	0.7606 / 2.7589	87.5%	-6.9%
large-mixed	300000	12.8945 / 179.9068	1.4041 / 5.7415	1.4815 / 13.2795	89.1%	-5.5%
many-blocks	297000	9.1445 / 14.5886	1.1627 / 3.8307	1.3062 / 4.6193	87.3%	-12.3%

P95 is noisy on this machine (one official large-mixed sample reached 179.9 ms, consistent with GC or system load). 60 samples per variant is a small sample; these are single-process, single-machine numbers and are not a model task, end-to-end latency, token usage, memory, or success-rate measurement. The pruner only runs when compaction conditions are met, so no overall task speedup is implied.

Web switches

The client half registers a conversation.hero.options slot entry with two buttons: Linear prune and Fast count (experimental). Selecting a state switches the session's agent preset through the official ctx.remote.agentPresets.select(sessionId, presetId) endpoint — no upstream code is modified or monkey-patched.

Three derived preset snapshots are registered from the current standard preset:

Preset id	enabled	fastLength
leet-comparison-off	official provider only	n/a
leet-comparison-linear	true	false
leet-comparison-fast	true	true

• standard and any custom presets are never overwritten; registration fails closed and rolls back on a collision.
• A preset is a startup snapshot. Buttons only act on the current blank session and on sessions created afterwards. A started session is rejected by the Host (agent-preset/locked); the UI disables the buttons and explains why. The plugin never claims an in-flight session switches instantly.
• fastLength is disabled in the UI until enabled is on, with the dependency stated in the button tooltip.
• Save failures, connection loss and server-side refusals are shown in a role="status" live region; a busy state blocks double submission. Buttons are native <button>s with aria-pressed, aria-describedby and a labelled group, so they work from the keyboard.
• On restart or reload the state is read back from the session's agentPreset projection, not from a client-only flag.

Lifecycle limitation (important): a session records the preset id it used. After this bundle is uninstalled, sessions that used leet-comparison-linear or leet-comparison-fast may no longer be able to recover that composition, because those snapshots name this package as their pruner provider. The leet-comparison-off snapshot uses only official providers and stays resolvable. Switch affected sessions back to standard before uninstalling. The preset descriptions state this too.

See VERIFICATION.md for exactly which Web checks were actually performed, on which DSH version, and which remain open.

Compatibility and installation

• DSH 0.2.2-rc.1 is the version actually used for parity tests. engines.dsh is >=0.2.2-rc.1 <0.3.0; other versions in that range were not verified.
• Node.js >= 24.9.0. Upstream reference commit fcc0ad462172522aaf5c93a7b2e01f6b9f0b0159.
• Peer dependencies must resolve to the running installation's copies: @deepseek-ai/dsh-compaction-tool-result-pruner 0.2.2-rc.1, @deepseek-ai/schemastery ~3.19.0.
• Pure JavaScript, no build step: the ./client export points at a checked-in browser factory script (a lazy-CJS window.__ModuleLoader__.load registration), and the root export is the Node plugin. Nothing is generated at install time, so a Git install needs no prepare script and no pnpm allowBuilds allowlist. Pin a commit anyway.

# Git source, pinned to a released tag (no build permission needed)
dsh plugin --profile YOUR_WEB_PROFILE add github:Harzva/dsh-leet-accelerator#v0.2.0
# ...or pin the exact commit SHA instead of the tag

# Or a released tarball
dsh plugin --profile YOUR_WEB_PROFILE add ./dsh-leet-accelerator-0.2.0.tgz

# Inspect the composed layer, then boot
dsh --profile YOUR_WEB_PROFILE --dump-config
dsh --profile YOUR_WEB_PROFILE

# Upgrade / remove
dsh plugin --profile YOUR_WEB_PROFILE update dsh-leet-accelerator
dsh plugin --profile YOUR_WEB_PROFILE remove dsh-leet-accelerator

--dump-config only proves the layer composition. It does not prove that a Web session actually uses the provider; see the verification record.

Switches outside the Web UI

Two headless ways to control the same two flags, without touching a user profile:

1. Bundle patch default — cordis.patch.yml inserts dsh-leet-accelerator with enabled: true. A later --patch overlay can override the row by id; remember a patch replaces the whole config, so restate thresholdChars, headChars, tailChars.
2. npm run control — a local HTTP page that writes control/accelerator.patch.json in the plugin directory only. Pass it with dsh --profile NAME --patch /absolute/path/to/control/accelerator.patch.json "task". It stores the setting read at the next startup, does not detect a running instance, and must never be exposed to a network or edited by two processes.

The bundle patch disables the official tool-result-pruner row in the same scope. Do not enable both providers for one toolResultPruner scope. Web sessions use isolated compaction groups inside their agent preset, which is why the preset-based switches above exist.

Tests and reproduction

npm ci
npm test          # 17 tests: four switch combinations, Unicode edges, session replay, client unit tests
npm run benchmark # writes benchmark/last-run.json

test/parity.test.js compares all four combinations against the official 0.2.2-rc.1 implementation: threshold boundaries, 300 seeded random multi-block Unicode cases, isolated surrogates, ZWJ sequences, odd head/tail budgets, invalid config, and full pruneSession event/replay parity. test/web-presets.test.js covers preset derivation plus client-side slot/state-machine unit tests with a fake React — those are unit tests, not browser or session-scope proof.

Market status

Per the official plugin index docs, DSH ships no built-in index source. Discover reads a third-party index.json that a user adds manually; indexes are third-party content. A GitHub dsh-plugin topic helps humans find the repository but does not put a plugin into any DSH market. This project is not listed in any concrete market index. See VERIFICATION.md for the format checks performed.

Known limits

• No real-model before/after comparison exists. The single real request returned HTTP 402 / QUOTA: Insufficient Balance; no tool ran and no pruner fired. No mock result is presented as a real one.
• fastLength regresses on emoji-heavy and multi-block inputs and is off by default.
• Preset-based switching cannot change a session that has already started.
• Derived preset ids can break historical session recovery after uninstall (see above).
• Parity is verified only against DSH 0.2.2-rc.1; parity with other releases in the declared range is untested.
• The top-level bundle patch does not control Web sessions' isolated compaction groups; use the derived presets for those.

License

MIT. Based on DeepSeek Harness, MIT, Copyright (c) 2026 DeepSeek. The upstream copyright and permission notice are retained in LICENSE.