dsh-github-issue

@zhangj1164/dsh-github-issue

English | 中文

GitHub issue generation and optimization service for DeepSeek Harness. Builds structured issue reports from telemetry analysis, constructs pre-filled issue-creation URLs, and optimizes natural-language issue descriptions with the configured model.

Plugin

GithubIssueService extends TypertRemoteService with static inject = ['llm']. The service uses ctx.get('llm') for model calls and exposes three Remote methods callable by the client UI.

Config · Default · Meaning repoUrl · https://github.com/zhangj1164/dsh-mega-plugins · Default repository URL for issue prefill when a request omits repoUrl. maxPrefillUrlLength · 7000 · Longest pre-filled issue URL to produce, counted in characters. GitHub refuses a request URL that is too long and shows an error page instead of the new-issue form, and it publishes no stable constant for the boundary, so this carries headroom and is deployment-tunable. 0 disables shortening. prefillTruncationNote · a Chinese note · Appended to the issue body when the URL had to be shortened, so the reader knows the body is partial. provider · unset · Provider route for model calls. Unset follows the deployment's agentDefaultModel selection, so a deployment can pin issue generation from cordis.yml alone. model · unset · Model id for model calls. Unset follows the same fallback as provider.

Remote methods

Method · Behavior generateReport(request) · Calls the model with a built-in structuring prompt to produce a uniform GitHub issue report from telemetry failure analysis. Returns a GithubIssueReport with title, body, and labels. The request names the pluginIds it covers — one report spans a whole suite — plus the analysis window and, per failure group, the route, the last failure time, and the attempts that followed it; the per-group facts are optional, and a missing one is passed to the model as not recorded rather than omitted. prefilledIssueUrl(request) · Builds a pre-filled GitHub issue-creation URL from a report. Validates the repo URL; returns invalid-url on failure. Shortens the body when the composed URL would exceed maxPrefillUrlLength, which is a budget the title and labels also spend and percent-encoding inflates, so the check is on the URL rather than the body alone. optimizeIssue(request) · Rewrites a natural-language description into a structured issue following a pinned Markdown template. Returns empty-input for blank descriptions, route-missing when no route resolves, llm-failure when the model call failed.

Model route

Both model-calling methods resolve their route the way the memo service does: the request's provider/model, then this service's Config, then the deployment's agentDefaultModel selection. The request fields are therefore optional, and the client UI sends them only when the user picked a model — the browser never has to publish a route for a call to work.

A blank value counts as absent rather than as a value, so an omitted field still reaches the deployment default instead of travelling to the model as an unusable route. When nothing resolves a route, the call returns route-missing and no model call is attempted: a call without an adapter cannot succeed, and reporting it as llm-failure would name the model for a routing problem.

Failure facts

A failed call keeps what DSH reported instead of collapsing it into one message. llm-failure carries the machine-routing code and the route in its message — model call to provider "custom" model "glm-5-2-260617" failed: NO_ADAPTER: no adapter registered for provider "custom" — so NO_ADAPTER (a route this deployment never registered), AUTH, RATE_LIMIT, and EMPTY_RESPONSE (the model genuinely returned no text) stay distinguishable. Reporting all of them as "the model produced no output" is what sent a user hunting for a model problem in a request that never reached a model.

A stream that throws a transport error rather than yielding a terminal finish chunk becomes LLM_STREAM_THREW, not an exception escaping the call.

Telemetry

Both model-calling methods report to the local telemetry service when the deployment has one, and behave identically when it does not. The service is read with ctx.get('telemetry') and deliberately not declared in static inject: this package installs on its own bundle, and a diagnostic dependency must never be the reason a service refuses to activate.

Event · Recorded A call that worked · Action optimizeIssue / generateReport as success, with the route that served it A failed call · The preserved DSH code, message, and HTTP status, the route, and the github-issue: feature anchor A call with no resolvable route · Code NO_MODEL_ROUTE — the same code @zhangj1164/dsh-memo uses for that condition, so one report can group both packages

generateReport also records which plugin's analysis it was given. Recording only successful calls would answer "which route served these calls" while leaving "has that route changed since it broke?" open, so both halves are recorded.

Issue report template

Both generateReport and optimizeIssue use built-in system prompts that pin a uniform Markdown structure. The first line is always a ## heading used as the issue title. The template includes sections for plugin, action, expected/actual behavior, route and timing, and reproduction steps.

generateReport receives the analysis window and, per failure group, the route, the last failure time, and how many attempts of the same action followed it. Its system prompt requires those facts to be stated, forbids proposing a mechanism the analysis does not evidence — sampling parameters, token budgets, timeouts, chunking, and response parsing are named as examples — and requires a fact the analysis does not carry to be written as not recorded. Writing that collection must be added is explicitly ruled out: a missing fact can be an older event as easily as a gap, and the report cannot tell which.

Requirement mapping

• Req 8 — generateReport turns telemetry failure analysis into a uniform GitHub issue report that correlates each failure group against the plugin's feature code.
• Req 9 — prefilledIssueUrl builds a pre-filled GitHub issue-creation URL so the user can jump to the project's new-issue page with the report already filled in.
• Req 10 — Packaged as a standalone DSH plugin: this package declares its own dsh.bundle with cordis.patch.yml, so a deployment can install it on its own and any UI plugin can reach it. See "Adopting this plugin" below.
• Req 11 — optimizeIssue rewrites a natural-language description into a structured issue using the configured model and a built-in structuring prompt. The ui-memo package calls this directly for its "Add Issue" editor.
• Req 12 — Reusable: the issue report template, the prefill URL builder, and the optimization prompt are all generic and not memo-specific.

Adopting this plugin

Install the package and add it to your deployment's composition:

pnpm add @zhangj1164/dsh-github-issue

# your cordis.yml, or your own bundle's patch
- insert:
    - id: github-issue
      name: @zhangj1164/dsh-github-issue
      config:
        repoUrl: https://github.com/your-org/your-repo

Read the service from any plugin with ctx.get('githubIssue'). It is optional from the consumer's side, so guard the read:

const issues = ctx.get('githubIssue')
if (issues !== undefined) {
  const { report } = await issues.generateReport({ analysis, repoUrl })
}

The service has no dependency on @zhangj1164/dsh-memo; memo is one consumer among possible others.

This bundle is the sole owner of the github-issue entry id. @zhangj1164/dsh-memo injects the service but does not insert this row, so the standalone bundle and the memo suite compose without a conflict and pnpm run verify-bundle-entries finds no overlap. A memo deployment therefore names the whole suite:

dsh plugin --profile web add @zhangj1164/dsh-telemetry @zhangj1164/dsh-github-issue @zhangj1164/dsh-memo @zhangj1164/dsh-client-ui-memo

@zhangj1164/dsh-github-issue is named explicitly rather than pulled in as a peer: a profile sets autoInstallPeers: false, so an install brings in only what the command names, and @zhangj1164/dsh-memo alone would leave this row unresolvable.

Exports

Every subpath resolves to the flat lib/*.js layout tsdown emits.

Subpath · Contents @zhangj1164/dsh-github-issue · The GithubIssueService plugin and its Config. @zhangj1164/dsh-github-issue/types · Wire types: GithubIssueReport, request and result shapes. @zhangj1164/dsh-github-issue/client · The browser-facing type face. @zhangj1164/dsh-github-issue/invariant · Invariant definitions.

Known Limitations

• Model dependency — inject: ['llm'] means the service will not activate without an LLM provider.
• Service-only surface — this package registers no model tool and ships no browser UI. Installing it on its own gives a deployment ctx.get('githubIssue') and nothing a user can see or click; the visible entry point is @zhangj1164/dsh-client-ui-memo's issue editor and log-analysis action, which reach it over Remote. Treat it as a backend that a UI plugin consumes, not as a standalone feature.