Each phase now carries a hash (Crockford Base32 identifier) and a
one-line title alongside the existing name/description/acceptance.
This gives agents immediate semantic context in the prompt without
needing to load full phase details from CAS.
Refs #23
The moderator now treats completedPhase matching the last planned phase name
as full completion. Also recognizes sentinel values (all-done, all_done,
complete) for robustness.
Fixes#21
小橘 🍊（NEKO Team）
RFC-001 was severely outdated (pre-refactor types). New architecture.md
covers the current three-phase engine, pure data roles, AgentBinding,
ExtractFn, and all design decisions.
小橘 <xiaoju@shazhou.work>
workflow-role-coder and workflow-role-planner have no test files —
bun test exits 1 on empty. Changed to 'echo no tests' for clean CI.
小橘 <xiaoju@shazhou.work>
dryRun no longer needed — tests use mock agents + mock fetch instead.
Removes isDryRun from WorkflowFnOptions, dryRun from ExtractConfig,
dryRunMeta from RoleDefinition, --dry-run from CLI, and all related
plumbing in engine/worker/fork/extract.
小橘 <xiaoju@shazhou.work>
BREAKING: Major architecture change.
- RoleDefinition = pure data (systemPrompt + schema + dryRunMeta)
- AgentFn = (ctx: ThreadContext) => Promise<string>, reads ctx.currentRole
- WorkflowDefinition decoupled from agents, bound via AgentBinding at runtime
- createWorkflow(def, binding, extract) replaces createRoleModerator
- Meta extraction moved into engine loop
- Delete workflow-util-role package (createRole, decorators, extract all gone)
- Role packages become pure data exports
- Agent packages updated to single-arg AgentFn
小橘 <xiaoju@shazhou.work>
- CreateRoleArgs.systemPrompt simplified to string (no more ctx callback)
- buildAgentPrompt injects real ctx.threadId in Tools section
- All four roles compute prompt at construction time from config only
- Removed duplicate ctx.threadId injection from reviewer/committer prompts
小橘 <xiaoju@shazhou.work>
- SolveIssueRolesConfig.agents allows per-role AgentFn overrides
- PlannerMeta now outputs phases (name, description, acceptance)
- CoderMeta reports completedPhase, works one phase at a time
- Moderator routes coder→coder until all phases done, then reviewer
小橘 <xiaoju@shazhou.work>
- Add RoleDefinition<Meta> = { description, run, schema } to core types
- WorkflowDefinition now carries description and RoleDefinition per role
- Add buildDescriptor() in core to derive WorkflowDescriptor from WorkflowDefinition
- Remove buildDescriptorFromRoles / RoleDescriptorInput from workflow-util-role
- Update solve-issue template, examples, and all tests
小橘 <xiaoju@shazhou.work>
WorkflowDefinition no longer carries 'name' — name is a registry-level
concern, not a bundle property. Also fixed all missing threadId in test
fixtures to match the updated ThreadContext type.
小橘 <xiaoju@shazhou.work>
CommitterMeta is now a 3-way discriminated union:
- committed: success with branch + commitSha
- recoverable: coder can fix (hook failures, lint, test, conflicts)
- unrecoverable: can't be fixed by code (auth, permissions, disk)
Moderator routes recoverable → coder for retry.
The package only contains createLlmAdapter (OpenAI chat → AgentFn),
which is an agent adapter, not a role. Aligns with workflow-agent-cursor
and workflow-agent-hermes naming.
- Add typecheck script (bunx tsc --build) to package.json
- Update check script to run typecheck before biome
- Fix mock fetch casts in test files (bun-types preconnect)
- Fix RequestInfo → Request | string | URL in llm-extract test
- Fix ThreadContext generic cast in solve-issue-template test
- Fix git-exec.ts missing return and module resolution
- Remove @types/node from workflow-role-committer
- Add exports field to workflow-util-agent/package.json
- Remove hardcoded git commands (git-exec.ts deleted)
- CommitterMeta is now a discriminated union: committed | failed
- Agent handles git operations, committer extracts structured result
- Moderator can route on meta.status for retry logic
Closes#17
- Add bun-types and @types/xxhashjs to root devDependencies
- Add types: ['bun-types'] to root and package tsconfigs
- Fix AST Node type narrowing in bundle-validator.ts with AcornNode type and narrowNode helper
- Fix generic ThreadContext variance in create-role-moderator.ts
- Add explicit parameter types in worker.ts
- Fix ChildProcess type in worker-spawn.ts to match spawn() stdio config
- Remove all tsconfig.tsbuildinfo from git tracking
- Add tsconfig.tsbuildinfo to .gitignore
Move pure role utilities (decorateRole, withDryRun, onFail, schemaDefaults)
into @uncaged/workflow-util-role. extractMetaOrThrow stays in role-llm
since it depends on LLM capabilities.
Dependency graph (no cycles):
util-role → workflow
role-llm → workflow, util-role
committer → workflow, util-role, role-llm
Closes#15
- Committer: git add/commit/push with LLM-generated branch+message
- Reviewer: code review role with approval meta
- Both use zod@4 schemas, no nerve-core deps
- 98 tests pass, biome clean
Closes#12
小橘 <xiaoju@shazhou.work>
- First param is now { prompt, steps } instead of bare prompt
- steps: [] for new thread, pre-filled for fork/resume
- createRoleModerator naturally handles resume via moderator routing
- No special replay logic needed
小橘 <xiaoju@shazhou.work>
Breaking design change:
- Section 2: WorkflowFn is now an AsyncGenerator that yields RoleOutput
- Section 8: Role/Moderator demoted from contract to optional helper
- Engine controls the loop via generator protocol (DIP)
- Zero injection — bundles don't import engine types
小橘 <xiaoju@shazhou.work>
- CLAUDE.md: functional-first, type over interface, no optional props, Result type
- biome.json: lint + format config (Biome 1.9)
- tsconfig.json: strict mode, composite project references
- .cursor/rules: global conventions + no-dynamic-import
Adapted from uncaged/nerve for bun workspace.
小橘 <xiaoju@shazhou.work>
xiaoju
xingyue