[rig-tasks] Daily rig evaluation — 2026-08-19 — 10/10 passed

[github-actions]

Summary

Task	Description	Typecheck	Key finding
1 (reused)	OS env variable scanner	✅ pass	Clean use of defineTool with s.enum literal return types; as const casts needed for union narrowing
2 (reused)	Sequential commit pipeline workflow	✅ pass	Good 3-agent sequential workflow with call() chaining; null guard after each call is idiomatic
3 (reused)	TOML config key extractor	✅ pass	s.record(s.record(s.string)) nested correctly; p.readInput correctly wired to input schema
4 (reused)	CSV column stats reporter	✅ pass	s.optional(s.number) used correctly for min/max/mean; handler return types well-typed with union
5 (reused)	Git worktree analyzer	✅ pass	addons: [steering(), repair()] ordering correct; s.enum on both type and status fields
6 (reused)	TypeScript narrowing pattern detector	✅ pass	Async defineTool handler with node:fs/promises works; p.glob for discovery pattern correct
7 (new)	Git reflog summarizer	✅ pass	p.bash with custom --format flag; s.enum with 6 action types; repair addon adds robustness
8 (new)	TypeScript export reporter with p.writeOutput	✅ pass	p.writeOutput used in instructions; s.path output field correctly declared
9 (new)	Parallel project health workflow	✅ pass	Promise.all with call.json steps; inlined p\...`in second call.json;s.int` for healthScore
10 (new)	Markdown frontmatter extractor	✅ pass	steering() + repair() both attached; extractFrontmatter tool returns hasFrontmatter boolean guard

---

Problems encountered

No failures this run. All 10 programs typechecked successfully on the first attempt.

---

Improvement opportunities

Missing or undiscoverable schema helpers (s.*)

• s.int vs s.number: The distinction is easy to miss — generated code sometimes reached for s.number for counts. Documenting this in a rule ("use s.int for all integer counts and line numbers") in a visible place would reduce drift.
• Nested s.record: s.record(s.record(s.string)) for TOML sections works but there's no example of nested records in SKILL.md. A short example of two-level records would help.

Missing or undiscoverable prompt helpers (p.*)

• p.writeOutput: Only appears in SKILL.md's table, not in any sample before this run. An example showing p.writeOutput in context (in instructions, and how the output schema maps to it) would make this more discoverable.
• p.readInput vs p.readAllInput: Programs that need to read a single file from input reliably reach for p.readInput, but the distinction from p.readAllInput (for an array) is not obvious without reading the references.

Error message quality

No typecheck failures observed. All errors seen in prior runs for similar patterns (e.g., wrong s.* for optional fields) were avoided by using the SKILL.md table.

API ergonomics

• defineTool handler return types with as const: When a handler returns a union ("path" | "locale" | ...), TypeScript widens the return to string unless as const is used. This is a recurring pattern in tool handlers returning s.enum values. A note in the defineTool reference about as const return casts would help.
• Promise.all vs parallel(): In workflows, Promise.all([call(...), call(...)]) works but SKILL.md also mentions parallel(). Both worked here — clarifying the preferred idiom (or confirming both are valid) would reduce ambiguity.
• Null guards after call(): Every sequential workflow step requires a if (!result) return null guard. This boilerplate is easy to forget. A lint rule or a helper would reduce noise.

Candidate lint rules

Rule: defineTool-enum-return-needs-as-const

• Invalid: return "path"; (inside a handler returning a union type)
• Valid: return "path" as const;
• Why model-confusing: TypeScript widens untagged string literals to string, causing runtime type mismatches vs declared s.enum output.
• Autofix possible: Yes — add as const to bare string literal returns in defineTool handlers.

Documentation gaps

• p.writeOutput needs at least one sample (now added: 438). The reference file prompt-intents.md should show the typical pattern: declare in instructions, map to an s.path output field.
• workflow body null guard pattern should be documented with a note explaining why it's needed (call() returns null on failure).

---

Tasks run today

• (reused) OS environment variable scanner with defineTool + s.enum categories
• (reused) Sequential workflow subagent pipeline (collect→classify→aggregate)
• (reused) TOML config file key extractor with defineTool + repair
• (reused) CSV column statistics reporter with defineTool + s.optional
• (reused) Git worktree listing analyzer with steering + repair addons
• (reused) TypeScript type narrowing pattern detector with async defineTool
• (new) Git reflog entry summarizer with s.enum action classification
• (new) TypeScript export reporter with p.writeOutput
• (new) Parallel project health workflow with call.json + Promise.all
• (new) Markdown frontmatter multi-field extractor with steering + repair

Generated by Daily Rig Task Generator · sonnet46 127.2 AIC · ⌖ 9.24 AIC · ⊞ 6.8K · ◷