DeepJIT

A JIT compiler plugin for deepseek-harness (dsh).

中文文档

Overview

DeepJIT watches agent execution traces, mines recurring "hot" workflows, and compiles them with an LLM into reusable skills (markdown) or flows (step templates) — then feeds them back into the running harness, no restart. It is for anyone who repeats similar multi-tool workflows and wants dsh to turn them into fast, reusable assets.

traces ──► SQLite ──► hot-path mining ──► LLM compile ──► skills / flows ──► dsh

• Artifacts live under ~/.dsh/deepjit/ and hot-reload into dsh.
• JIT never compiles its own tools, so it can't feed on itself.
• Lifecycle (compiler-style): an AOT pass validates flows against the live tool registry and constant-folds literal args; tiering promotes hot, reliable skills to flows and deoptimizes unreliable flows; GC prunes stale artifacts.

Compatibility

Item	Value
DSH version	@deepseek-ai/dsh 0.1.5-rc.2 (runtime-verified)
DSH mainline	session-persistence drill-down now uses SessionHandle.open().read(offset,length) (0.1.5 API); summarizer supports both the legacy readFrom and the new handle form
Node	^22.19 || >=24
Profiles	headless, web

dsh is pre-release; APIs may drift. Pins @deepseek-ai/* to 0.1.5-rc.2 (cordis 4.0.2).

Install / Uninstall

# install from npm (published as @fly3366/deepjit)
dsh plugin --profile web add @fly3366/deepjit

# or install from git (no npm release needed)
dsh plugin --profile web add github:fly3366/DeepJIT

# disable for one profile
dsh plugin --profile web remove @fly3366/deepjit

# fully remove local data
rm -rf ~/.dsh/deepjit

Quick start

dsh plugin --profile headless add github:fly3366/DeepJIT
DEEPSEEK_API_KEY=... dsh --profile headless "read package.json and tsconfig.json, then summarize"
# repeat similar tasks; deepjit mines and compiles hot flows automatically
dsh --profile headless "use deepjit_status to list compiled artifacts"

Configuration

Override in cordis.patch.yml or a profile patch. Key options (full list in src/config.ts):

Key	Default	Description
enabled	true	master switch
summarizeIntervalMs	600000	JIT cycle (mine + compile)
minRepeat	3	min occurrences for a hot sequence
argumentAware	false	include sorted arg-key signatures in mined sequences
compileCandidates	1	best-of-N candidates per compile; higher = better quality, more LLM calls
minFlowSteps	2	min tool steps a flow must have to be compiled
minPatternValue	6	min value score (count × steps) to justify a compile
flushBatchSize	200	trace rows per batched SQLite write
maxPendingCalls	10000	cap for in-memory pending/raw maps (bounds memory)
minerMaxRows	20000	max trace rows read per session per mining cycle
transcriptMaxRows	2000	max tool rows read per compile transcript
gcEnabled / gcStaleMs / gcProtectMs	true / 14d / 1d	GC: disable artifacts unused beyond gcStaleMs after a gcProtectMs grace
dryRun	false	publish artifacts as disabled; enable manually via deepjit_status
traceRetentionMs / patternRetentionMs	7d / 7d	prune trace rows / stale uncompiled patterns older than this
deoptMinUses / deoptMaxSuccessRate	5 / 0.5	disable a flow used ≥N times with success rate ≤ this (deoptimization)
qualityMinUses / minQuality	5 / 0	disable active artifacts used ≥N times whose quality score < minQuality (0 = off)
promoteMinUses / promoteMinSuccessRate	5 / 0.8	recompile a hot, reliable skill's pattern as a flow (promotion)
llmProvider / llmModel	deepseek-official / (session)	compile model; empty = reuse session model
locale	auto	en / zh / auto (dsh locale → LANG → English)

Sensitive: no keys are stored. The compile call uses dsh's credential service or the launching environment's DEEPSEEK_API_KEY.

Permissions & data

• Files: writes only under ~/.dsh/deepjit/ (SQLite traces, skills, flows, log); reads session JSONL via ctx.sessionPersistence for compile drill-down.
• Network: LLM calls go through dsh's ctx.llm (DeepSeek provider); no other network access.
• Credentials: none stored; resolved by dsh or the environment.
• User data: stores compact execution traces (tool args/results, message text).
• Tools: flow steps run through ctx.tools.execute and the normal permission gates.
• Observability: dsh's OTel telemetry covers agent sessions only; DeepJIT keeps its own counters (traces flushed, compiles, LLM latency, GC/deopt/promote) and emits GenAI semantic-convention spans (gen_ai.*: system/model/usage tokens) for each LLM call. Read counters via deepjit_status {action:"metrics"}; spans export via dsh-o11y-plugin when present.

Troubleshooting

• Log: ~/.dsh/deepjit/deepjit.log. Database: ~/.dsh/deepjit/deepjit.db.
• MISSING_CREDENTIAL → export DEEPSEEK_API_KEY or store it in dsh's Models page.
• TRANSPORT/NO_ADAPTER on compile → usually a transient LLM call failure; deepjit retries and falls back to the next cycle.
• Roll back: dsh plugin --profile <p> remove deepjit, then rm -rf ~/.dsh/deepjit.

Development

npm install && npm test     # node:test, run natively via Node type stripping
npm run typecheck && npm run build

See CONTRIBUTING.md and AGENTS.md.

Related work

The "compile recurring workflows just-in-time" idea and the best-of-N compile selection are inspired by JIT-Agent — Scaling Harness Intelligence via Just-in-Time Harness Evolution (bingreeky/JIT). DeepJIT adapts that harness-evolution concept to run online inside a live dsh session, compiling recurring tool flows into skills/flows rather than synthesizing whole harnesses.

License & security

MIT. Report vulnerabilities privately per SECURITY.md.