roadmap: deterministic incident evidence → runtime adapters

[duhenri9]

Canonical roadmap

This issue replaces the temporary Debug Evidence checkpoint that lived in wm3-request-protocol before this standalone repository existed.

Product thesis

Diagnose before you mutate: collect runtime evidence, preserve contradictory signals and produce a reproducible debug bundle before proposing code changes.

V0 — deterministic incident bundle

Build one offline incident fixture with one real root cause and plausible decoys.

Required mechanisms:

• structured ingestion for logs, stack trace, environment metadata and recent-change evidence;
• secret/environment redaction before persistence;
• bounded hypothesis model;
• evidence links classified as support, contradiction or unresolved;
• allowlisted safe probes only;
• SUPPORTED | CONTRADICTED | INDETERMINATE semantics;
• deterministic input/bundle digests;
• no source mutation in V0.

Required controls:

1. misleading-log control — loudest error is not root cause;
2. missing-evidence control — return INDETERMINATE rather than certainty;
3. unsafe-probe control — policy blocks a proposed destructive/unapproved probe.

V0.2 — real local runtime inputs

• Python and Node log/trace parsers;
• local Git diff/recent-commit correlation;
• runtime/environment fingerprint with allowlisted values only;
• safe filesystem/process metadata adapters;
• reproducible incident archive.

V0.3 — observability adapters

Only after the local evidence contract is stable:

• OpenTelemetry traces;
• structured JSON logs;
• Sentry-compatible import where practical;
• container/Kubernetes read-only diagnostic adapters.

V0.4 — optional model reasoning

• provider-neutral hypothesis adapter;
• redaction before model calls;
• model suggestions remain untrusted hypotheses;
• deterministic evidence engine owns result semantics;
• prompt/model/version identity recorded.

V0.5 — remediation proposals

Proposal only first. Mutation requires a separately designed authority contract and must not be smuggled into the diagnosis layer.

Public-quality gate

• one-command offline demo;
• positive + all three negative controls;
• tests/static checks;
• pinned/minimal-permission CI;
• evidence artifact uploaded by CI;
• privacy/redaction threat model;
• architecture/ADR docs;
• no paid model key required;
• benchmark bundle-generation latency/resource use before performance claims.

Non-goals

• autonomous code repair in V0;
• arbitrary shell agent;
• replacing an observability platform;
• presenting LLM explanation as evidence;
• ingesting private/commercial project data as public fixtures;
• production incident auto-remediation.