dsh-lint-loop
The change-aware quality loop for DeepSeek Harness.
Fix what the agent broke. Ignore what was already broken. Prove the change is clean.
A lint wrapper tells the agent what is broken. Quality Loop compares this turn against evidence captured before the first edit, repairs lint regressions safely, and checks types and affected tests before completion. Historical failures stay visible without demanding an unrelated cleanup. A structured Quality Receipt says exactly what ran, what changed and what could not be verified.
v0.6.0 — Quality Loop
English | 中文
Before → after
Before the task Agent changes src/parser.ts
old eslint warning introduces a fixable lint error
old TS2322 in legacy.ts breaks parser's return type
old failing migration test makes parser.test.ts fail
Ordinary lint wrapper Quality Loop
shows old + new lint debt Fast Lane: new lint error → safe repair
misses type/test regressions Completion Lane: lint delta + types + impacted tests
can say done after lint passes steers only for NEW regressions, at most twice
keeps the old failures as historical debt
emits the verification scope and final verdict
After the agent fixes the new type/test regressions:
Quality Receipt: clean · 1 changed · lint clean / typecheck clean / tests clean
Ignored historical debt: old lint, old type error, old test failure
Tests executed: parser.test.ts; unrelated tests excluded when the dependency is provable
clean means no new regressions in the executed scope, not “the whole repository passes.” Missing baselines, unsupported runners, skipped tests, discovery limits, malformed output and timeouts produce incomplete, never a green claim. The default gate steers only on attributable new regressions; incomplete evidence remains visible in the receipt.
Two speeds
Fast Lane: the existing edit → lint delta → prompt feedback loop. lint_repair safely repairs this turn's lint regressions. It never repairs type/test failures by rewriting code. File fixers skip historical fixable debt and roll back newly introduced lint findings or failed verification.
Completion Lane: at the awaited agent/turn-stopping checkpoint, or via quality_verify, check all this turn's files against the original baselines. Capture Node/TS typecheck and test baselines once before the first edit; this initial capture may run package suites and has a separate bounded budget. Later edits retain the original evidence. A continuation retains the entire turn scope, even without another edit event.
There are at most two forced continuations and two autofix rounds per turn by default. Exhausting either budget does not turn a regression receipt green. Reads and failed writes do not join the change set. No source rewinding, stash or worktree reset is used to obtain a baseline.
Install
Install the npm package into your DSH profile:
dsh plugin --profile web add dsh-lint-loop@0.6.0
For local development:
pnpm install --frozen-lockfile
pnpm build
dsh plugin --profile web add /absolute/path/to/dsh-lint-loop
No linter, compiler or test framework is bundled: use the repository's installed tools. Node 22/24 are supported. The DSH peer policy remains >=0.1.0-rc.1 <0.2.0-0 || >=0.2.0-0 <0.3.0-0; the existing Cordis peer remains ^4.0.1.
Modes
Use balanced unless you have a reason to change it. Override the existing lint-loop plugin row in your profile patch:
- id: lint-loop
config:
mode: balanced
| Preset | Fast Lane | Completion Lane | Sweep budget |
|---|
fast | lint delta + on-demand safe repair | lint only; types/tests explicitly skipped | 60s |
balanced (default) | lint delta + on-demand safe repair | lint delta, repository package typechecks, impacted tests with conservative fallback | 60s |
strict | same safe repair rules | lint delta, repository package typechecks and repository package test suites | 120s |
fast deliberately yields an incomplete full-quality receipt. strict broadens test scope; it still ignores historical failures and never forces unlimited work. Advanced options override the preset.
| Option | Default | Meaning |
|---|
completionChecks | preset | enable typecheck/tests |
qualityTimeoutMs | preset | total budget for each baseline, completion or repair sweep |
qualityMaxFiles | 2000 | discovery entry budget, including directories |
qualityMaxChecks | 32 | maximum package typecheck/test attempts per sweep |
gate, gateMaxSteers, gateSeverity | true, 2, error | bounded continuation and lint severity |
autoInject, settleMs, sectionTtlMs, sectionSeverity | true, 600, 30000, error | post-edit prompt feedback |
linters, linterPath | auto, {} | existing linter selection and executable overrides |
timeoutMs, maxFindings | 10000, 50 | per-linter timeout and displayed finding cap |
codeFrames, frameLines, frameLimit | true, 1, 5 | compact source frames |
Existing configuration names and tool arguments are retained. Package-scoped Go/Rust diagnostics keep their timeout floors; Completion Lane's overall cancellation budget still applies.
Tools and Quality Receipt
| Tool | Use |
|---|
quality_verify | Read-only Completion Lane; returns the complete structured Quality Receipt |
quality_receipt | Retrieve this session/repository's last receipt without running checks |
lint_status | Current turn's lint regressions and ignored debt |
lint_repair | Bounded, regression-only safe autofix; call after editing |
lint_diagnostics | File findings; scope: all / introduced / preexisting, severity and max filters |
lint_workspace_errors | Errors in files checked this session, with regression scope |
lint_fix | One-file repair; an existing baseline enforces regression-only scope |
quality_verify and quality_receipt expose the structured receipt in ordinary DSH tool results. Pure Host presenters describe running / clean / regression / failed, changed-file counts and check statuses for consumers that use that contract. A desktop/Web Quality Bar is deferred: the audited DSH 0.2 desktop client's renderer does not consume Host presentCall / presentResult. Delivering a dedicated bar needs client integration; v0.6 does not add a client bundle or UI hack. The automatic checkpoint saves evidence; quality_receipt retrieves it.
UI audit and future capture guide · Actual desktop Receipt
Impacted tests
v0.6 supports local Vitest (test: vitest / vitest run) and Jest (test: jest / jest --runInBand) with their JSON reporters. Typechecks run local TypeScript via tsc --noEmit --pretty false --incremental false -p tsconfig.json for every discovered package scope. Project references are marked unavailable because --noEmit cannot prove rebuilt reference outputs.
Selection follows transitive reverse relative imports, including re-exports, .js → .ts resolution, index modules and directly changed *.test.*, *.spec.* / __tests__ files. An isolated source with no provable test edge falls back to its package. Unresolved/bare workspace imports, dynamic dependencies, configuration/setup/fixture changes and customized test discovery fall back to repository package suites. Monorepo checks execute in each package's cwd and can use hoisted local binaries.
Fallbacks are recorded as package-fallback or repository-fallback with reasons. A fallback runs the wider supported suites; a discovery budget overflow stays incomplete even if discovered tests pass. Custom/chained scripts, node:test and non-Node test frameworks are reported unavailable instead of running an arbitrary command or claiming coverage. ImpactProvider.select(plan, changed) is the optional integration seam for a future code-index provider; v0.6 has no dependency on dsh-code-index.
Regression semantics and receipts
Lint uses the existing source-aware matcher. Typecheck compares multisets of file + TS code + normalized message + source context, ignoring line shifts. Tests compare suite path + full test name + failure signature; a changed failure in an already-failing test is a regression. Startup/suite-load failures remain incomplete when they cannot be attributed reliably. Resolved failures count as agentFixed only after successful checks; deleted/skipped tests cannot be advertised as fixes.
Receipts include changed files, new regressions, auto-fixed and agent-fixed issues, ignored historical debt, three check statuses, actual commands/arguments and executed test identities, baseline evidence, selection reasons, autofix/continuation rounds, elapsed time and final verdict. They are isolated per session/repository. Turn completion, errors and cancellation reset baselines and repair budgets, cancel pending checks and retain the last receipt until plugin disposal. A read returns a detached snapshot. Native tool results carry their receipt through DSH's ordinary durable replay path.
Automatic repair cannot broaden file scope. eslint/biome/ruff retain safe file repair. Go/Rust package/crate repair is skipped because source snapshots cannot guarantee protection for manifests, lockfiles or neighbors. Their diagnostics and explicit standalone fixer APIs remain available. Calling lint_fix on a file with no earlier baseline is the backwards-compatible explicit debt-cleanup path; use lint_repair for agent-turn repair.
Where this fits
| Plugin | Primary purpose |
|---|
| dsh-lint-loop | Prove this turn did not introduce lint/type/test regressions; ignore historical debt |
| dsh-doublecheck | Requirements interrogation and delivery discipline, including adversarial review |
| dsh-test-runner | General structured test execution and failure summaries |
| dsh-review-loop | Human incremental diff review and checkpoints |
No LLM code review, requirements management, security scan, coverage platform or CI platform. No required companion plugins and no new linter families.
Limits
- Only successful DSH
edit/write tool mutations and the compatible fs-intent path are authoritative. Shell/custom-tool edits, deletions and concurrent external mutations are not fully attributed. A receipt describes the tracked set, not every git diff.
- Initial test/type baselines cost one bounded repository-package sweep. Flaky tests can look like regressions; suites may have side effects. Checks use local processes, as the existing linter runner does; they are not a sandbox boundary.
- Static dependency selection is conservative, not a complete language/runtime index. Generated files, assets and nonliteral runtime dependencies require broader evidence; unknown patterns should be configured to a supported runner or assessed separately.
- Missing linters, missing compiler/test runners, custom scripts, TypeScript reference builds, skipped tests and unparsed output remain incomplete. The gate does not demand historical cleanup or invent a result.
- Host presenters are tested against DSH 0.1/0.2 contracts. The current built-in desktop/Web renderer does not consume them; a dedicated Quality Bar and its screenshot/GIF are deferred.
Architecture audit and design · Candidate validation evidence
Development
pnpm install --frozen-lockfile
pnpm typecheck
pnpm test
pnpm build
CI covers Windows/Linux × Node 22/24. Tests include real TypeScript/Vitest integration, historical debt, duplicate diagnostics, safe autofix/rollback, impacted selection/fallback, monorepos, bounded retries, cancellation/timeouts, native presenters, disposal/remount and Windows argument handling.
Feedback and discussions · MIT License
