[rig-tasks] Daily rig evaluation — 2026-07-24 — 5/5 passed

[github-actions]

Summary

Task	Description	Typecheck	Key finding
1	Dead code detector (reused: deadcod6)	✅ pass	s.optional(s.int) for line numbers works cleanly
2	Git blame hotspot analyzer (reused: githtsp1)	✅ pass	s.record(s.object(...)) for file map works but needed a wrapper files field — root s.record output shape works
3	Code complexity scorer with subagents (reused: cmplxsc2)	✅ pass	p.readInput("filePath") required for dynamic paths in subagents — not obvious from docs
4	Build log analyzer with repair (new: bldlog3y)	✅ pass	repair addon mentioned but not exported from "rig" — users must implement custom AgentAddon
5	TypeScript type alias mapper with defineTool (new: tstype4z)	✅ pass	defineTool with s.object parameters + s.record(s.object) root output composes cleanly

Problems encountered

repair and steering are documented but not exported

The SKILL.md references repair and steering as first-class addons:

"Add a repair addon with maxTurns: 3"

But neither is exported from "rig". The built-in repair behavior is triggered by maxTurns on the spec itself, not via an addon. When a generator (or user) tries:

import { agent, repair, s } from "rig"; // type error: 'repair' is not exported

The correct idiom is just maxTurns: 3 in the spec, but this is never stated explicitly alongside the addon documentation. Users are left confused about what addons are for vs. maxTurns.

Reproduction sketch:

// What docs imply is possible:
const a = agent({ addons: [repair()], maxTurns: 3, ... });
// Actual: repair is not a named export; maxTurns on spec is the repair knob
const a = agent({ maxTurns: 3, ... }); // correct

rig-expander subagent produced no output

When used as a background agent in this evaluation workflow, the rig-expander subagent completed (status: idle) but returned empty turn content. All 5 programs had to be written manually. This may be a runtime/transport issue with background agent invocation, but surfacing this as a known limitation or adding a health-check mechanism would help.

p.readInput discoverability

When generating code for a subagent that accepts a file path as input, the natural instinct is to use p.read(input.path) — which is a type error since p.read requires a literal string. p.readInput("fieldName") is the correct approach but is documented in a subsection that's easy to miss. A brief note in the agent({ input }) docs would help.

Improvement opportunities

Missing schema helpers (s.*)

• s.path — a string schema variant with semantic meaning "file path"; would improve readability and could hint to the runtime about path-based context resolution.
• s.int in record values — s.record(s.int) is documented but s.record(s.object(...)) as a root output schema lacks an explicit example showing how to give the whole agent output a record shape. An example in the Schema section would help.

Missing prompt helpers (p.*)

• p.repair(customPrompt) — an exported addon factory that wraps the built-in repair logic with a user-supplied fallback message. Currently the user must write a full AgentAddon function to customize repair prompts.
• p.steering(fn) — an exported addon factory for steering (mentioned in SKILL.md but not exported). Even a thin wrapper would be more ergonomic than raw AgentAddon.

Error message quality

• When addons receives a non-function value, the error is "Agent addon entries must be functions." — this is clear, but doesn't say which addon index failed or what value was passed. Including the value's type/shape would help debug mistakes like addons: [repair] where repair is undefined.

API ergonomics

• maxTurns vs repair addon: The relationship is confusing. maxTurns controls built-in repair retries, but addons exists for custom logic. The docs should explicitly state: "For built-in repair, set maxTurns in the spec. The addons array is for custom turn-level middleware — not for configuring built-in repair."
• Root-level s.record(...) output: Works correctly, but many patterns want a mixed output with both a record of results and a summary string. This forces a wrapper s.object({ files: s.record(...), summary: s.string }). A note acknowledging this pattern would help.
• Subagent agents field naming: Subagent keys in agents: { fileAnalyzer } must match the variable name exactly. There's no runtime validation error if the key is wrong — the harness silently ignores it. A check or clearer docs on naming conventions would reduce confusion.

Documentation gaps

• p.readInput placement: Currently in a dedicated subsection far from the p.read docs. Should be mentioned inline next to p.read with a note: "Use p.readInput(field) when the path is supplied by the caller as an input field."
• Built-in repair behavior: The section on maxTurns should clarify that repair (re-prompting on JSON/schema failure) is automatic when maxTurns > 1 — no addon needed.
• tools field in agent(spec): The tools field in AgentSpec is not mentioned in SKILL.md at all. The defineTool section jumps straight to passing tools via agent({ tools: [...] }) without listing tools in the spec fields table.

Tasks run today

• (reused) Task 1: Dead code detector that finds potentially unused exported symbols using p.bash grep and s.optional fields in output schema
• (reused) Task 2: Git blame hotspot analyzer using p.bash git log with stat, s.record output keyed by file path with changeCount/lastAuthor/risk s.enum, repair addon maxTurns:3
• (reused) Task 3: Code complexity scorer with subagent delegation: coordinator uses p.bash find + p.read per file via fileAnalyzer subagent, aggregates scores with s.enum verdict
• (new) Task 4: Build log analyzer using p.bash npm run build, repair addon maxTurns:3, outputs s.object with s.array of error objects having s.enum severity (error/warning/info), s.optional file, and s.boolean buildSucceeded flag
• (new) Task 5: TypeScript type alias mapper using p.bash find + file content scan, defineTool for regex-based categorization, outputs s.record(s.object) with s.enum kind (primitive/union/intersection/mapped/other) and s.boolean exported per alias

Generated by Daily Rig Task Generator · sonnet46 59.6 AIC · ⌖ 7.5 AIC · ⊞ 5.5K · ◷