- workflow-generator, sense-generator, gitea-issue-solver all now have:
planner → coder → reviewer → tester → committer → END
- Reviewer uses createHermesRole with git diff/status for static analysis
- Checks: garbage files, secrets, debug code, unrelated changes
- Planner prompt now requires Role Behavior sections for every role
- Coder prompt now emphasizes reading initial user prompt for specifics
Hermes agent cwd is not necessarily the nerve workspace root.
Inject nerveRoot into tester prompts so all file paths and commands
use absolute paths to the workspace directory.
- Both workflow/sense tester: build first, then validate, then list
- workflow-generator: remove nerve workflow dry-run (doesn't exist)
- sense-generator: add build step before sense list/trigger
- done=true means all files created, build+lint pass
- done=false means progress made, continue next iteration
- Updated both sense-generator and workflow-generator coder prompts
- Move index.js → src/index.ts with proper types for all 4 senses
- Move schema.ts → src/schema.ts
- Add package.json with esbuild build script per sense
- Bundle to index.js at sense root (daemon loads this)
- Update sense-generator coder prompt with TypeScript conventions
Fixes#224
- Add esbuild build script to package.json
- Add .gitignore for dist/
- Remove stale tsc-generated .js files (now bundled)
- Include sense files generated by workflow run (hermes-gateway-health update, worker-process-metrics)
- Clean up nerve.yaml (remove deleted workflows, migrate reflexes to interval)
- Remove nerveYaml injection from planner (skill has it)
- Remove sensesDir/nerveRoot from coder and tester (skill has conventions)
- Prompts now just say 'read the skill' instead of inlining knowledge
- BuildSenseGeneratorDeps reduced to { provider, cwd }
- index.ts drops getNerveYaml(), SENSES_DIR, readFileSync
Every role is self-contained (types.ts, prompt.ts, index.ts).
No shared.ts, no cross-role imports. All dependencies injected:
index.ts — wiring (resolve env, call buildSenseGenerator)
build.ts — buildSenseGenerator(deps) → WorkflowDefinition
moderator.ts — pure routing, composes meta from role types
roles/planner/ — buildPlannerRole(deps), self-contained
roles/coder/ — buildCoderRole(deps), self-contained
roles/tester/ — buildTesterRole(deps), self-contained
Workflow is now reusable: buildSenseGenerator() can be called with
any provider/paths, not hardcoded to this machine.
小橘 🍊（NEKO Team）
attempt is a deterministic fact (count of tester steps), not something
an LLM should guess. Removed from tester meta schema, moderator counts
it directly from context.steps.
小橘 🍊（NEKO Team）
Replaced 122-line hand-written smoke test with createHermesRole.
Hermes runs the full lifecycle check autonomously via terminal:
files → nerve.yaml → sense list → trigger → query → logs on failure.
All 3 roles now use factories:
planner: createCursorRole (ask mode)
coder: createCursorRole (default mode)
tester: createHermesRole (e2e verification)
小橘 🍊（NEKO Team）
- planner/coder: replaced 80+ lines hand-written agent calls with createCursorRole()
- SenseMeta slimmed to routing signals only (senseName, filesCreated, passed/attempt)
- Roles read context from thread via nerve thread <id>, not from previous role's meta
- tester stays hand-written (pure CLI logic)
- Re-exported spawnSafe from workflow-utils for helper use
Refs uncaged/nerve#210
小橘 🍊（NEKO Team）
llmExtract returns {} as T in dryRun mode, causing .map() on undefined.
Each role now returns mock data immediately when dryRun is true.
小橘 🍊（NEKO Team）
