[rig-tasks] Daily rig evaluation — 2026-08-30 — 10/10 passed

[github-actions]

Summary

Task	Description	Type	Typecheck	Key finding
1 (reused)	Dotenv vs process.env drift detector	agent	✅ pass	s.record + defineTool + repair() composed cleanly; as const needed on return literals
2 (reused)	Vitest snapshot count workflow	workflow	✅ pass	workflow() meta requires description field — omitting it caused initial failure; Promise.all + explicit type annotations on reduce needed
3 (reused)	Git stale branch reporter	agent	✅ pass	defineTool returning s.enum literal with as const worked correctly
4 (reused)	TypeScript abstract class finder	agent	✅ pass	p.glob + p.bash combined cleanly; s.boolean output field from tool worked well
5 (reused)	Dockerfile ENV instruction inspector	agent	✅ pass	p.bash find chained correctly; tool returned structured object without issue
6 (reused)	NPM script command prefix analyzer	agent	✅ pass	p.read + s.record(s.object(...)) + s.record(s.int) — smooth pattern
7 (new)	TypeScript default parameter extractor	agent	✅ pass	p.glob + p.bash + repair() addon; implicit any on reduce avoided with explicit annotations
8 (new)	Git branch risk classifier	agent	✅ pass	steering() + repair() addon ordering ([steering(), repair()]) required attention
9 (new)	Test file coverage pipeline	workflow	✅ pass	3-stage sequential call() workflow; call() returns `T
10 (new)	JSON schema type distribution	agent	✅ pass	defineTool with JSON.parse in handler; s.record(s.int) for type distribution map

---

Problems encountered

No typecheck failures this run after fixing two issues discovered during development:

1. workflow() meta missing description (task 2): The initial version used meta: { name: "..." } without description. TypeScript surfaced: Property 'description' is missing in type '{ name: string; }' but required in type 'WorkflowMeta'. Fix: add description: "..." to all workflow meta objects.

2. workflow body destructuring and call type (task 2): Initial version used body(call) but the correct form is body: async ({ call }) => .... Also, Promise.all + call() results are typed T | null, requiring null-coalescing when accessing fields.

3. reduce implicit any (task 2): .reduce((sum, f) => sum + f.lineCount, 0) failed with implicit any errors. Fix: explicit type annotations (sum: number, f: { lineCount: number }).

---

Improvement opportunities

Missing or undiscoverable schema helpers (s.*)

No gaps found. s.record(s.int) for count maps and s.record(s.object(...)) for keyed maps worked as documented. The SKILL.md table covers these well.

Missing or undiscoverable prompt helpers (p.*)

• p.glob → p.bash find combo is commonly needed when a glob finds paths but content also needs to be fetched. A p.readGlob(pattern) helper that inlines file content would reduce this pattern.
• p.readOptional was referenced in several pool descriptions but not exercised; it is documented in SKILL.md but the fallback parameter semantics could use an inline example.

Error message quality

• The WorkflowMeta.description is missing TypeScript error is clear and actionable.
• The call type error (This expression is not callable. Type 'WorkflowContext<unknown>' has no call signatures) was confusing — the actual issue was destructuring body(call) instead of body: async ({ call }) =>. A targeted lint rule or better error message could help here.

API ergonomics

• call() return type is T | null: Every caller must null-coalesce. This is correct behavior on agent failure, but it adds boilerplate at every call site in workflows. A helper like call.required(...) that throws on null might improve workflow readability.
• as const on tool return literals: Required for enum schema comparison, but easy to forget. A lint rule catching return "value" in a defineTool handler where the return type is s.enum(...) would catch this reliably.
• steering() + repair() ordering: The rule is [steering(), repair()] but the SKILL.md table says "Retry with final-turn warning" without making the required order visually prominent. A note in the addons section would help.

Candidate lint rules

Rule: no-untyped-reduce-in-workflow

• Invalid: .reduce((sum, f) => sum + f.x, 0)
• Valid: .reduce((sum: number, f: { x: number }) => sum + f.x, 0)
• Why: TypeScript cannot infer accumulator types in .reduce() when the input comes from a call() result typed as T | null; this is a recurring pattern in workflow bodies.
• Autofix: safe (inject : number and matching element type based on the array item schema).

Rule: require-workflow-meta-description

• Invalid: workflow({ meta: { name: "x" }, body: ... })
• Valid: workflow({ meta: { name: "x", description: "..." }, body: ... })
• Why: description is required in WorkflowMeta but the compile error only appears at the call site, not at the meta literal. A dedicated lint error points at the missing field directly.
• Autofix: not safe (description content must be human-written).

Documentation gaps

• SKILL.md "High-frequency decisions" table does not mention workflow() body destructuring syntax ({ call, phase, pipeline }). A one-liner example would prevent the body(call) confusion.
• The addons ordering rule for [steering(), repair()] is stated in SKILL.md but buried in prose; a table row in the "High-frequency decisions" section would make it scannable.

---

Tasks run today

• (reused) dotenv vs process.env drift detector: reads .env.example, p.bash grep process.env usages, defineTool + repair, s.record output
• (reused) Vitest snapshot count reporter workflow: two subagents via Promise.all — snapshot scanner + parser
• (reused) Git stale branch age reporter: p.bash git branch -r, defineTool classifyBranchAge with s.enum return
• (reused) TypeScript abstract class finder: p.glob + p.bash, defineTool extractAbstractClass, s.boolean output
• (reused) Dockerfile ENV instruction parser: p.bash find Dockerfiles, defineTool parseEnvInstruction
• (reused) NPM script command prefix analyzer: p.read package.json, defineTool classifyScriptPrefix, s.record outputs
• (new) TypeScript default parameter extractor: p.glob + p.bash + defineTool + repair() addon
• (new) Git branch risk classifier: p.bash + defineTool + steering() + repair() in order
• (new) Test file coverage pipeline: 3-stage sequential workflow() + call() pipeline
• (new) JSON schema type distribution: p.glob + p.bash + defineTool + s.record(s.int) aggregation

Generated by Daily Rig Task Generator · sonnet46 198.2 AIC · ⌖ 8.45 AIC · ⊞ 6.8K · ◷