try-works/role-model / .agents/skills/recursive-mode
recursive-mode
UnverifiedRepository workflow orchestration skill for staged implementation, locked artifacts, late-phase receipts, and durable memory maintenance. Use when executing recursive-mode runs, resuming a run, locking a phase, or verifying locks.
references/benchmark-addon/recursive-benchmark/references/patterns.md
View on GitHub# recursive-benchmark patterns
## Table of Contents
- [Purpose](#purpose)
- [Paired Benchmark Rule](#paired-benchmark-rule)
- [Runner Selection](#runner-selection)
- [Project Fixture Rule](#project-fixture-rule)
- [Scenario Tier Rule](#scenario-tier-rule)
- [Execution Rule](#execution-rule)
- [Logging Rule](#logging-rule)
- [Scoring Rule](#scoring-rule)
- [Report Rule](#report-rule)
- [Failure Handling](#failure-handling)
## Purpose
Use this reference when the user wants a concrete benchmark that compares recursive-mode against a non-recursive baseline.
## Paired Benchmark Rule
Always create two comparable arms:
- `recursive-off`
- `recursive-on`
Both arms must start from the same benchmark project spec. The recursive-on arm may add recursive-mode scaffolding and a run-local `00-requirements.md`, but it must not receive a different product scope.
The recursive-off arm should receive the same requirements and other controller guidance in the prompt only, not as repository or workspace benchmark docs.
## Runner Selection
Record the exact runtime used for each benchmark:
- runner name
- provider family
- model string
- timeout budget
If the runtime does not expose token usage, report that as unavailable instead of guessing.
## Project Fixture Rule
Packaged benchmark fixtures should be:
- realistic enough to require multi-file implementation work
- local-browser friendly
- runnable from a temp folder
- free of database or external service requirements
## Scenario Tier Rule
Support packaged tiers such as:
- easy
- medium
- hard
- xhard
The harness should let the user choose a scenario explicitly instead of hardcoding one benchmark forever.
## Execution Rule
Use disposable repos and bounded timeouts.
Recommended first-version behavior:
1. copy the packaged benchmark starter repo into temp repos; the scaffold may be bootstrap-only and should not be assumed to contain starter product code
2. provide the same benchmark requirements to both arms, while keeping recursive-off requirements in prompt text only
3. bootstrap the live recursive scaffold only in the recursive-on repo, wrap the benchmark requirements into a recursive-compliant run-local `00-requirements.md`, and prompt the agent to read `/.recursive/RECURSIVE.md` plus the bridge docs before starting
4. invoke the selected coding-agent CLI non-interactively
5. run a mandatory controller-side judge review for every completed arm, preferring `gpt-5.4` and falling back to the benchmarked model when necessary
6. update per-arm progress files so status can be inferred from file changes in the benchmark workspace
7. evaluate the finished repos with build, test, preview, recursive-run-artifact checks, and controller-side recursive lint for the recursive-on arm; for Rust nested worktrees, use an isolated evaluation snapshot when parent repo manifests would otherwise interfere
If the runtime and provider can handle it, the harness may also offer a parallel arm mode. If a runner proves unstable in parallel, the harness should automatically downgrade that runner to sequential execution and note the downgrade in the report.
## Logging Rule
Capture:
- raw runner output
- setup and execution timestamps
- build/test/preview command results
- screenshot artifact paths
- timeout or failure details
- in-repo agent activity log path
- hint log path when hints were provided
- per-arm progress artifact path
- report path
When the in-repo activity log is incomplete, the harness may fall back to controller-visible file timestamps and should report that it did so.
## Scoring Rule
Use a fixed rubric that rewards:
- successful build
- successful tests
- successful local preview
- major feature coverage
- presence of benchmark logging/evidence
The harness report should distinguish:
- runner or CLI health
- product outcome
- heuristic harness score
- any later controller-side review score
- combined benchmark score that blends heuristic coverage and judge review
- recursive-on worktree isolation status sourced from `00-worktree.md`
- recursive-on closeout/lint status and any repo-root delivery drift that violates the declared product root
Recommended default weighting:
- heuristic harness score: 70%
- mandatory judge metric: 30%
Keep the rubric transparent and packaged with the fixture.
## Report Rule
The final report should behave like a scoreboard:
- summary comparison table
- per-arm score breakdown
- recursive workflow completion status for the recursive-on arm
- recursive-on worktree isolation status and recorded worktree location
- mandatory judge metric with reviewer provenance
- live progress artifact paths
- screenshot sections with image embeds when available
- notable failures and missing evidence
- artifact and log paths
- timestamp fallback evidence when used
## Failure Handling
If one or both arms fail:
- keep the logs
- keep the score breakdown
- clearly mark timeout, crash, or incomplete implementation state
- still write the report instead of aborting without results
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 468k | Updated | Claude CodeCodex | Unverified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 319k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 319k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 319k | Updated | Claude CodeCodex | Verified | Install |