DeepSeek Harness plugin porting pi-advisor-flow: an on-demand executor/advisor workflow with consultation tool, review gates, usage ledger, privacy tiers and secret redaction.
The plugin will be installed here. Keep web if you are unsure.
npx -y @deepseek-ai/dsh plugin --profile web add github:ccll/dsh-advisor-flow#95dc6663c9ca0a3457e146d08e55d6d948e78c76
Compatibility and provenance
Advisor Flow is published as dsh-advisor-flow and currently resolves to version 0.2.1. The Hub verifies its manifest and preserves the exact installation source for reproducible installs.
A DeepSeek Harness (DSH) plugin that ports the pi-advisor-flow executor/advisor workflow to DSH. Everyday work runs on the everyday model; at key decisions, after repeated failures and before a turn closes, a stronger advisor model provides a second opinion. The advisor only advises — it never takes over the work (no tools, no file edits), and the session owner keeps full configurability and visibility over cost, privacy and blocking behavior.
Observable behavior is aligned with pi-advisor-flow 0.8.2 (divergences are listed item by item in RATIONALE.md).
On-demand consultation: the executor calls the ask_advisor tool whenever it wants a second opinion — with no arguments for a general review, or with a question, a draft, a repo-context tier (it can only narrow the owner's allowed tier) and a hand-picked file list. Every answer carries an adviceId for later reference.
Manual consultation: the session owner can trigger a consultation at any time with /advisor-manual [focus], and cancel one in flight with /advisor cancel.
Hard loop gate: when the same tool call with an equivalent signature repeats up to the threshold (default 3), the host intercepts the call before it executes and asks the advisor to review; the advisor answers with a three-value decision — proceed | revise | blocked — and blocked is handled according to the configured blocking mode (warn-and-continue | block-tool | block-session).
Executor guidelines: before-plan / after-failure / before-completion guidelines are injected into the system prompt as soft constraints (on by default; the text matches the pi 0.8.2 English original verbatim).
Hard completion review (default mode): in the default hard intervention mode, every executor turn is reviewed by the advisor before it closes; a revise or blocked verdict sends the full advice back to the executor to keep working. In soft mode the plugin behaves exactly like pi-advisor-flow: guidelines only.
Outcome recording: the record_advisor_outcome tool lets the executor voluntarily record whether an advice was followed and validated (one-shot JSONL entries, advice text stored as an HMAC hash; off by default).
Usage ledger and status: per-call and cumulative input/output/cache tokens with cost details, categorized by trigger (on-demand / manual / gate); /advisor status shows the model routing, gate state and gate decision statistics.
Privacy tiers: repo context off | summary | full, file contents never leave the machine without an independent opt-in (ownership-checked), and secret-shaped values are replaced with placeholders before anything is sent — redaction runs before truncation.
Non-blocking guarantee: any advisor or gate failure never stalls the main agent loop.
Install
dsh plugin --profile web add dsh-advisor-flow
The npm package ships prebuilt, so no local build step is needed. If the settings card does not appear after installing, restart dsh web once. A listing on dsh-market (dsh plugin --profile web add dshmarket), the plugin market inside DSH, is in progress via the curated awesome-dsh-plugin index.
npm publication is in progress — until it lands, install straight from this repository:
dsh plugin --profile web add github:ccll/dsh-advisor-flow
No host source patches and no postinstall scripts — plugin integration is declared entirely through the standard dsh.bundle manifest (cordis.patch.yml declares the plugin's own insertion, not a modification of DSH). Version assumptions about DSH's plugin seams are declared in package.json under dsh.compat.
Requirements
DSH web, tested against @deepseek-ai/dsh@0.1.5-rc.1 (see dsh.compat in package.json).
An advisor model reachable through DSH's own model routing — pick any provider/model already configured in DSH; the plugin reuses the host LLM service and never talks to providers directly.
Configuration
Configuration lives in the advisor-flow namespace of settings.yaml and can also be edited in the web settings card. Changes apply live, no restart needed. Highlights:
Enable prerequisites: advisor-flow.enabled: true plus advisor.provider / advisor.model; with an incomplete configuration the whole feature stays disabled and the reason is queryable via /advisor status.
mode — intervention strength: hard (default; forces a completion review before every turn closes) or soft (guidelines only, pi behavior).
advisor.callTimeoutMs — overall timeout for one consultation, default 600000 ms (10 minutes).
advisor.maxTokens — output cap for one consultation; by default it follows the host's configured cap for the selected model.
Loop-gate threshold defaults to 3 (minimum 2); the three guidelines and the loop gate are on by default; secret redaction is off by default (enable it in the settings card — recommended, secret-shaped values are then sent as placeholders).
/advisor-manual [focus] — start a manual consultation immediately; cancel one in flight with /advisor cancel.
/advisor [on|off|toggle|status|gates|cancel] — session-level switch and status queries; bare /advisor toggles.
Development
npm test # unit tests (node --test)
npm run build:client # rebuild the web settings-card client bundle
node scripts/screenshot.mjs shots # regenerate README screenshots in an isolated demo environment
Commit and verification gates (AgentMap living docs + .githooks/) are described in CONVENTIONS.md.
Acknowledgments & Disclaimers
This project ports the executor/advisor workflow of the MIT-licensed pi-advisor-flow (version 0.8.2) to DeepSeek Harness; the guideline and protocol texts are kept verbatim where the host allows it. Thanks to the original author for the design.
Advisor reviews are model output, not ground truth: treat them as a second opinion, not as a verification oracle.
99.99% of this project's code and documentation was written and reviewed by AI, so bugs and doc/code drift are quite likely. If you run into any problems, please open an issue.
