<div align="center">

<img src="https://raw.githubusercontent.com/kk3ya03-star/dsh-lcx-codex/main/assets/dsh-lcx-codex-banner.jpg" alt="DSH-LCX-CODEX" width="100%" />

DSH-LCX-CODEX

OpenAI Responses / Codex-native capabilities for DeepSeek Harness.

 !Node !DSH !License

简体中文 · English

</div>

---

dsh-lcx-codex is a community DSH plugin for GPT routes already configured through llm-pi-ai / openai-responses. It keeps DSH's Agent, Web, Session and Compaction ownership intact and fills the provider-native gaps.

> LCX is only the project name. This project is not affiliated with OpenAI, DeepSeek, Sub2API or NewAPI.

• Current DSH image pipeline: Native V2 reuses readImageRequest() and the active route image budgets, so compaction and ordinary GPT requests see the same deterministic image variants.

Features

rc.8: DSH 0.1.1-rc.2 native alignment

rc.8 targets DSH 0.1.1-rc.2 directly. Ordinary Hosted Search still follows the active Agent route, while Native V2 image serialization now uses DSH's readImageRequest() request-version pipeline instead of reading attachment master bytes.

Search routing still crosses DSH's tools/execute → SearchProvider boundary without changing the model-visible web_search schema:

Agent: lcx / gpt-5.6-luna
        └─ DSH web_search
             └─ Hosted Search: lcx / gpt-5.6-luna

The settings-page model is now a fallback used only when no active Agent route is available.

Hosted Search also receives its own stable cache namespace:

conversation replay: dsh-lcx:<route hash>
hosted search:       dsh-lcx-search:<route hash>

A search row with a low cache hit in NewAPI therefore does not imply conversation truncation.

Native V2 Compaction

DSH continues to own pressure, range selection, pruning, durable replacement, /compact, and overflow recovery. The plugin replaces only the purpose: compaction summarizer transport with Native V2:

DSH compaction transaction
        └─ purpose=compaction
              └─ /responses + compaction_trigger

Native success means zero basic-summary request. Basic summary runs only if Native fails and fallback is enabled.

Conversation fidelity layer

Provider-native opaque compaction is intentionally lossy. Real long-session tests showed that assistant-only details can disappear even when the opaque checkpoint itself is valid. The current checkpoint therefore keeps:

bounded client-visible history
+ bounded assistant-visible final answers
+ opaque Native V2 compaction item

The explicit prefix is bounded at roughly 64k estimated tokens, with up to about 24k reserved for assistant answers and about 3k per retained assistant answer. Reasoning, raw tool traces, large search payloads and telemetry are not copied back into the prompt.

The goal is not byte-for-byte replay. It is drop process noise, retain conversation facts.

Automatic compaction

0% -------------------- 90% ----- 95% ----- 100%
        normal             Native   emergency   hard cap
                           V2       DSH prune

Default behavior:

• below 90%: suppress stock 80% pressure pruning for compatible Native sessions;
• 90–95%: Native V2 first;
• 95%+: DSH's replay-safe tool-result pruning may run as an emergency guard;
• provider-confirmed context overflow keeps DSH's normal recovery path;
• manual /compact is unaffected.

Cache behavior

A provider cache miss is not the same thing as DSH deleting history.

Observed real-world pattern:

turn N:   ~155k uncached, cacheRead 0
turn N+1: ~1k new input, ~155k cacheRead

That means the full context was resent and cached again. Prefix-changing operations include compaction, emergency pruning, route/model changes, and other DSH surface replacement. Restart/idle/provider cache TTL may also force a cold request independently of this plugin.

Search modes

Ordinary web_search

Recommended default. It follows the current Agent GPT Responses route.

websearch_gpt_advanced

Opt in only for Hosted-only parameters such as domain filters, approximate location, search-context size and image search. Keeping it disabled avoids unnecessary tool-catalog churn.

websearch_alpha

Stateful Codex/Alpha-style search/open/find/click/screenshot; registered only after a matching capability probe.

Installation

Stable channel:

dsh plugin --profile web add dsh-lcx-codex

Pre-release channel (current rc.8):

dsh plugin --profile web add dsh-lcx-codex@next

Local RC:

dsh plugin --profile web remove dsh-lcx-codex
dsh plugin --profile web add .\dsh-lcx-codex-0.4.0-rc.8.tgz
dsh web

Do not delete existing sessions or $DSH_HOME/storages/lcx-codex/ just to upgrade. The legacy v3 sidecar remains read-only compatibility state.

Recommended settings

Enable plugin                         on
GPT Hosted Search                     on
Advanced Hosted Search               off
Alpha Search                         off
Native V2 remote compaction          on
Native-first auto compaction         on
Native threshold                     90%
Emergency DSH prune                  95%
web_search timeout                   240s

Requirements

• Node.js >= 20
• DSH 0.1.1-rc.2
• a working GPT openai-responses route in DSH
• an upstream that actually supports the enabled Hosted Search / Native V2 / Alpha capabilities

Typical paths include direct Sub2API and NewAPI-relayed Responses routes.

Release channels

Stable releases use npm's latest tag. 0.4.0-rc.* builds publish to npm's next tag. GitHub tags must match package.json; Trusted Publishing runs tests before publishing.

Development

npm test
npm run test:schema
npm pack --ignore-scripts

See [ARCHITECTURE.md](ARCHITECTURE.md) and [CHANGELOG.md](CHANGELOG.md) for implementation details.

License

MIT