dsh-codex-bridge

A DeepSeek Harness bundle that lets the agent read the local Codex CLI quota and delegate hard tasks to codex exec with a model and reasoning effort that fit the current ChatGPT-plan budget.

It is a host-only DSH plugin: two tools, no browser half, no build step, and no npm dependencies.

Unofficial community plugin. Not affiliated with or endorsed by OpenAI, Codex, or DeepSeek.

What it adds

Tool	What it does	Codex usage
codex_status	Reads the 5-hour and weekly ChatGPT-plan windows, reset times, credits, the model ladder, and the quota-aware recommendation.	none
codex_delegate	Runs one non-interactive codex exec task in a chosen directory and returns the answer, token usage, thread id, and the quota left afterwards.	yes

codex_status talks to codex app-server (account/rateLimits/read, model/list). codex_delegate is the only path that spends model usage.

Quota policy

The deployment default is intentionally narrow:

• Default: gpt-6.1-sol at reasoning effort high.
• Reserve: when either window has 10% or less left, auto switches to gpt-5.6-sol at medium.
• gpt-6-astra: never selected automatically. An explicit model: "gpt-6-astra" is refused unless the caller also sets allow_expensive_model: true, because it consumes the shared quota much faster.
• Floors: the model never goes below gpt-5.6-sol; reasoning effort never goes below medium.
• Blocked: at 95% used, or when the backend disallows ordinary usage, runs are refused unless allow_low_quota: true.

Recommendation table (model=auto, reasoning_effort=auto):

Complexity	Normal (both windows > 10% left)	Reserve (either window ≤ 10% left)
simple	gpt-6.1-sol / high	gpt-5.6-sol / medium
standard	gpt-6.1-sol / high	gpt-5.6-sol / medium
hard	gpt-6.1-sol / high	gpt-5.6-sol / medium
critical	gpt-6.1-sol / high	gpt-5.6-sol / high

Explicit choices are validated, never silently downgraded. A refusal names the selectable models or the missing opt-in.

About gpt-6.1-sol

The installed Codex CLI build may not list gpt-6.1-sol in model/list even though the backend accepts it (a real codex exec run confirmed this during development). codex_status therefore shows it as a configured model and warns that codex exec may print a missing-model-metadata warning. Change defaultModel / minModel in the config if your Codex account uses different ids.

Install

In the DSH Web GUI, open Plugins, add the package directory or install the bundle through the plugin manager:

{ "action": "install_bundle", "target": "/path/to/dsh-codex-bridge" }

From GitHub (once the repository is published):

dsh plugin --profile web add github:AK-blank/dsh-codex-bridge#v1.1.1

From a local checkout:

dsh plugin --profile web add /path/to/dsh-codex-bridge
dsh --profile web --dump-config | grep -A2 codex-bridge

Then restart dsh web if the host process already cached an older module. Host-side plugin code is cached per process; a new bundle row activates live, but edits to lib/*.js need a restart.

Requirements: DSH 0.2.0-rc.2 (the tested release), Node.js 20+, and a signed-in Codex CLI (codex login).

Configuration

Override the row in the profile's cordis.patch.yml:

- id: codex-bridge
  name: dsh-codex-bridge
  config:
    minModel: gpt-5.6-sol
    defaultModel: gpt-6.1-sol
    expensiveModel: gpt-6-astra
    minEffort: medium
    defaultEffort: high
    reservePercent: 10
    stopPercent: 95
    statusCacheSeconds: 30
    defaultSandbox: read-only
    ephemeral: true
    maxTimeoutSeconds: 1800
    maxOutputChars: 16000
    codexCommand: codex
    codexHome: ''

Key	Default	Meaning
minModel	gpt-5.6-sol	model floor and reserve fallback
defaultModel	gpt-6.1-sol	normal default (may be configured even when not listed by the CLI)
expensiveModel	gpt-6-astra	explicit opt-in only
minEffort	medium	lowest reasoning level ever used
defaultEffort	high	reasoning level for ordinary and hard work
reservePercent	10	fall back when either window has this much or less left
stopPercent	95	block ordinary runs at this used percentage
statusCacheSeconds	30	quota snapshot cache
defaultSandbox	read-only	Codex sandbox unless the caller overrides it
ephemeral	true	do not persist each delegation under ~/.codex/sessions
maxTimeoutSeconds	1800	hard wall-clock limit per delegation
maxOutputChars	16000	answer characters returned to the model
codexCommand	codex	executable name or absolute path
codexHome	empty	empty inherits the host environment

Usage

Ask the agent:

Check the Codex quota, then delegate this hard problem to Codex with the right model and effort.

The agent calls codex_status, picks a complexity (or passes explicit model / reasoning_effort), then calls codex_delegate. The task must be self-contained: Codex cannot see the DSH conversation, so include absolute paths, symptoms, constraints, and the exact deliverable.

Default sandbox is read-only. Pass sandbox: "workspace-write" and a cwd only when Codex must edit files.

Verification

node selftest.mjs             # offline: package contract, policy, rendering, registration
node selftest.mjs --live      # also read the real quota and model ladder (no model usage)
node selftest.mjs --live-exec # also run one tiny codex exec task (spends a little usage)

The offline suite covers the quota tiers, the 10% reserve switch, the astra guard, the model/effort floors, blocked-quota refusal, rendering, and the bundle/registration contract. Live checks verify the real codex app-server response and the configured ladder. The repository's GitHub Actions workflow runs the offline suite on every push.

Limitations

• The app-server methods belong to the installed Codex CLI version; an incompatible upgrade fails with CODEX_RPC / CODEX_PROTOCOL instead of inventing data.
• Codex does not share this conversation's context.
• Default read-only means Codex can read cwd but cannot write; workspace-write keeps writes under cwd and still follows Codex's own sandbox/network policy.
• The returned quota is a percentage snapshot, not an exact token reservation; the tiers are spend guardrails.
• gpt-6.1-sol availability and model metadata depend on the Codex CLI build and account; configure other ids if needed.

Community listing

docs/awesome-entry.yml is the prepared entry for the community directory awesome-dsh-plugin (category usage). That directory requires the repository to be at least one day old before a PR is accepted, so the PR is intentionally not opened on publish day. The repository also carries the dsh-plugin topic for automatic indexes.

License

MIT. See LICENSE.