flow-trace-audit

flow-trace-audit

End-to-end reachability audit for any codebase. Finds every feature that exists in code but is not wired end to end: dead endpoints, broken calls, buttons that do nothing, responses nobody handles.

Tracing the flow is the source of truth.

Built and signed by AVAA DEV — L'agent IA gouverné. NO CODE NO DONE.

The problem

AI-assisted development generates capable modules fast — and quietly leaves them unwired. The code compiles, the tests pass, the button renders. And nothing happens when you click it. Backend commands sit unreachable; frontend calls target handlers that were renamed three sessions ago; error branches swallow failures silently.

No linter catches this. Type checkers don't cross the UI/backend boundary. It is the single most common failure mode of AI-generated applications.

What this plugin does

One skill, one method, any stack:

1. Phase 0 — Discovery: identifies the project's real communication mechanisms from the code (Tauri invoke, WebSocket, HTTP routes, IPC, gRPC, custom bridges…). No assumed architecture.
2. Phases 1–3 — Three exhaustive inventories: backend entry points, frontend emitted calls, UI trigger chains — each item proven with file:line evidence.
3. Phase 4 — Reachability matrix: cross-references the inventories and classifies every gap into nine statuses: BROKEN-CALL, INERT-TRIGGER, SWALLOWED-ERROR, DEAD-ENDPOINT, ORPHAN-RESPONSE, DEFINED-NOT-REGISTERED, PIPELINE-BREAK, CONFIG-DRIFT, COGNITION-BREAK.
4. Phase 5 — Règle #30: every command output is read with the five-pass protocol before any conclusion.
5. Phase 6 — Report with integrity check: the audit is valid only when every inventoried item is accounted for. Counts must balance.

Flux G traces the cognition of AI agents themselves — perception, real assembled context, recall, traced decision, action, feedback, consolidation — so an agent that ignores instructions, forgets, loops or claims work it did not do is diagnosed like any other broken flow.

The audit modifies zero project files. It produces a diagnostic you can trust, then you decide what to wire.

Install

/plugin marketplace add charlesphilippebelhadj/avaa-dev-flow-trace
/plugin install flow-trace-audit@avaa-dev

Use

Explicitly:

/flow-trace-audit:audit
/flow-trace-audit:audit src/features/settings

Or just ask naturally — the skill triggers on requests like "audit the wiring", "is everything connected end to end?", "why does this button do nothing?", "find dead endpoints".

Method provenance

The canonical methodology — Debug Pilleur Master, Universal Debugging and Tracing Methodology (universal version 2026-05-24, flux G 2026-10-26) — is bundled verbatim in skills/flow-trace-audit/references/DEBUG_PILLEUR_MASTER.md (canonical, French) with a full English translation in DEBUG_PILLEUR_MASTER_EN.md: 7 universal flux types (including agent cognition) with step-by-step verification chains, checklists and schemas, reference commands, 25 documented failure patterns, the Règle #30 five-pass output-reading protocol, the forbidden anti-patterns of AI-assisted debugging, and how to apply the method to development tasks, not only diagnostics.

It was forged on real projects at AVAA DEV, where a governed multi-agent delivery system enforces one doctrine: a feature is not done until its complete flow — trigger → transport → handler → effect → response — is traced and proven. Assumptions are forbidden; evidence is mandatory.

License

MIT — see LICENSE.