Core CLI
Protocol and verify commands expose stable JSON interfaces alongside doctor, demo, and benchmark operations.
← Work / Active files
[ CASE FILE / MIGHTY-MOUSE ]
ActiveMighty Mouse
A provider-agnostic reliability harness that gives AI coding agents structured protocols and project-native verification.
Mighty MouseA provider-agnostic reliability harness that gives AI coding agents structured protocols and project-native verification.6/10 vs 6/10 / First-attempt passes
01
The messy workflow
Agent output is not proof that a coding task is complete.
Coding agents can produce plausible explanations and still miss required files, violate scope, or stop before the project’s own checks pass. The operational problem is turning a model response into a bounded, inspectable result another person can trust.
Mighty Mouse supplies versioned complexity protocols, project-native verification, structured results, and a bounded retry contract. The model and agent platform remain replaceable; the reliability layer stays explicit.
02
Ownership boundary
I built the reliability layer, not the models or agent platforms.
I designed and built the harness, versioned protocols, project verifier, result schema, CLI, MCP transport, packaging, platform rules, CI, evaluation fixtures, and prospective study.
Foundation models, model runtimes, MCP itself, and platforms such as Codex and Antigravity are third-party systems. Mighty Mouse integrates with them; it does not claim their underlying work.
Declared ownership
Original reliability harness, versioned protocols, project verifier, evaluation design, packaging, CLI, MCP transport, platform rules, CI, and study execution. Foundation models, model runtimes, MCP itself, and agent platforms are third-party systems.
03
System map
How the work moves.
- Input
Task + complexity input
A coding task enters with an explicit complexity level or a deterministic default.
- Decision
Protocol selection
The harness returns a versioned low, medium, or high-complexity execution contract.
- State
Agent execution
The chosen agent works inside declared scope using the project’s existing tools and conventions.
- Decision
Project-native verification
Tests, lint, build, scope, and changed-file checks run as structured verification steps.
- Decision
Bounded retry loop
Failures return specific next steps; the agent may retry without entering an unbounded correction cycle.
- Output
Structured result
Pass state, checks, warnings, and suggestions are returned in a stable human- and machine-readable form.
- Failure path
Stop condition
A passing result completes the task; exhausted retries or unresolved scope failures stop and escalate.
04
Hard calls
Decisions that changed the system.
05
Build record
Selected implementation record.
MCP server
A separate package exposes protocol and verification tools over stdio for compatible agent platforms.
Installable packages
Core and MCP wheels and source distributions are built, archive-inspected, checksummed, and smoke-tested outside the checkout.
CI and paired evaluation
Python 3.10–3.13 CI, packaging gates, frozen trial manifests, and blind reviews keep implementation and claims auditable.
06
Verification ledger
Claims with their edges left on.
HS-7VERIFIED RECORD
6/10 vs 6/10
First-attempt passes
Mighty Mouse and the control each passed six of ten paired real-project tasks on the first attempt, with zero scope violations in both conditions.
- Sample
- 10 paired tasks producing 20 condition runs
- Boundary
- Ten paired tasks from the recorded projects, agent, models, and environment only; no generalized reliability or speed improvement was demonstrated.
Sourcedata/evidence/real_project_report.md — completed prospective study
LimitThe study did not demonstrate better generalized first-pass reliability.
HS-7VERIFIED RECORD
4 vs 6
Retry rounds
Mighty Mouse used four retry rounds across the sample; the control used six.
- Sample
- 10 paired tasks producing 20 condition runs
- Boundary
- Ten paired tasks from the recorded projects, agent, models, and environment only; no generalized reliability or speed improvement was demonstrated.
Sourcedata/evidence/real_project_report.md — completed prospective study
LimitTwo fewer retries in this sample is not evidence of a universal improvement.
HS-7VERIFIED RECORD
4.60 vs 4.30
Mean blind-review quality
Mighty Mouse received a 4.60 mean blind-review score versus 4.30 for the control across the ten paired tasks.
- Sample
- 10 paired tasks producing 20 condition runs
- Boundary
- Ten paired tasks from the recorded projects, agent, models, and environment only; no generalized reliability or speed improvement was demonstrated.
Sourcedata/evidence/real_project_report.md — completed prospective study
LimitThe quality difference belongs to this sample and its recorded blind-review procedure.
HS-7VERIFIED RECORD
262.5s vs 229.5s
Median duration
Mighty Mouse was slower: 262.5 seconds median duration versus 229.5 seconds for the control; raw mean duration was also slower.
- Sample
- 10 paired tasks producing 20 condition runs
- Boundary
- Ten paired tasks from the recorded projects, agent, models, and environment only; no generalized reliability or speed improvement was demonstrated.
Sourcedata/evidence/real_project_report.md — completed prospective study
LimitThe study did not demonstrate a generalized speed improvement.
HS-7VERIFIED RECORD
29.5%
Historical synthetic latency result
Lean reduced average latency by 29.5% while both conditions passed 15 of 15 tasks in the historical synthetic promotion suite.
- Sample
- 15 paired tasks producing 30 condition runs
- Boundary
- Historical synthetic promotion suite and recorded environment only; the later prospective real-project study did not reproduce a generalized speed or first-pass reliability improvement.
Sourcedata/evidence/results/PROMOTION_NOTES.md — historical promotion validation
LimitThis historical synthetic result is bounded to that suite and must not be generalized to real project work.
07
Visible proof
The system, without the sales pitch.
08
Failure + iteration
What failed stays in the record.
resolved
The synthetic suite hit a ceiling
- Observed
- Mighty Mouse, the historical protocol conditions, and a bare control all passed the frozen synthetic tasks, limiting what the suite could prove about reliability.
- Response
- Bound the 29.5% latency result to its recorded environment and ran a prospective paired study on real project tasks.
monitoring
The real-project result was mixed
- Observed
- First-attempt passes tied at 6/10, and Mighty Mouse was slower by both mean and median duration despite fewer retries and higher mean blind-review quality.
- Response
- Published the unfavorable timing and tied reliability results and stated that no generalized improvement was demonstrated.
resolved
macOS metadata contaminated package inputs
- Observed
- AppleDouble resource-fork files created noisy Git and archive failures on the external volume.
- Response
- Added package-data exclusions, source-distribution filtering, archive inspection, and explicit regression tests.
resolved
Editable installs and Python 3.10 exposed packaging gaps
- Observed
- Custom build-backend behavior and older-Python compatibility failed outside the original development environment.
- Response
- Completed editable-build hooks, added compatibility dependencies, and required clean Python 3.10–3.13 CI plus artifact-only installation smokes.
09
Current status
Released reliability tooling with bounded evidence.
Version 0.2.1 ships the core library, CLI, MCP transport, platform rules, verifier, CI matrix, release artifacts, and the completed ten-pair study. The current evidence supports fewer retries and higher mean blind-review quality in this sample, not a generalized first-pass reliability or speed improvement.
Open to the right work
A useful system starts with the actual problem.
Full-time product, automation, and technical-operations roles are the priority. Select workflow and creative-technology projects are open.