Audit Remediation Prioritizer

AGENT ROLE: Autonomous remediation-planning agent (capstone). Read findings from other audits, normalise + dedupe + score them, sequence a roadmap, build ticket payloads, and emit the JSON in references/output.schema.json. NEVER file tickets without explicit user approval — the script only produces dry-run payloads.

OBJECTIVE

Convert a pile of raw, overlapping audit findings into a defensible, sequenced plan: one scored record per unique issue, grouped into work templates, ordered into phases by return-on-effort, each with an owner, an estimate, and a filable ticket.

INPUTS

• findings (REQUIRED): comma-separated JSON files. Each may be a full audit output (Site Migration, Indexation Coverage, Index Bloat, Canonicalization) or a generic list/{findings:[...]} of {type, url|urls, severity, evidence, clicks}.
• traffic (OPTIONAL): JSON map url -> clicks|sessions to weight reach when a finding carries no traffic of its own.
• effort_map / owner_map / impact_map (OPTIONAL): JSON overrides finding_type -> value for the built-in tables.
• model (OPTIONAL, rice|ice, default rice), phases (OPTIONAL, default 3), jira_export (OPTIONAL flag): build Jira-shaped payloads (still dry-run).

AUTHENTICATION (none required)

• Scoring needs no credentials. Reach uses the traffic carried inside findings (e.g. Site Migration clicks_at_risk, Index Bloat clicks) or the optional --traffic map.
• Jira export is a PAYLOAD BUILDER only. This skill NEVER POSTs a ticket; export_mode is always dry_run. Filing tickets is a side-effecting action the agent must confirm with the user and perform explicitly outside this script. jira_credentials_present merely reports whether JIRA_BASE_URL/JIRA_TOKEN are set.

EXPECTED TOOL CALLS

• Run scripts/remediation_prioritizer.py --findings a.json,b.json[,...] [--traffic clicks.json] [--model rice] [--phases 3].
• The script auto-detects each file's shape (at_risk+redirect_class → migration; gaps+cause → indexation; results+action → bloat; conflicts+conflict_types → canonicalization; else generic) and maps every item to {finding_type, urls, traffic_at_stake, evidence, source}.

PROCEDURE (deterministic)

STEP 1 — NORMALISE every input item to the canonical finding model; carry each source's own traffic signal. STEP 2 — DEDUPE on (finding_type, sorted(urls)): merge duplicates across sources, keep the max traffic, union the source list, count merges. STEP 3 — SCORE each finding. impact (0.5–3) and effort (points) come from the built-in finding_type table (overridable); a severity hint of high/low clamps impact. confidence = 0.5 + 0.3 (has traffic) + 0.2 (known type) − 0.1 (no URLs), clamped [0.3,1.0]. reach = traffic-at-stake, or the affected-URL count when no traffic exists. RICE = reach·impact·confidence / effort; ICE = impact₁₀·confidence₁₀·ease₁₀ / 100. STEP 4 — GROUP findings into finding_type templates; aggregate score, URL count, and traffic; sort templates by score then traffic. STEP 5 — SEQUENCE templates into phases equal-rank bands (phase 1 = highest ROI). Build one ticket payload per template. EMIT summary, roadmap, per-finding scores, and tickets.

RATE LIMITS & ERROR HANDLING

• Pure local computation; no network, no rate limits. IF a --findings file is missing THEN STOP error.code="INPUT_FILE_MISSING"; IF it is not valid JSON THEN STOP error.code="INPUT_INVALID".
• Unrecognised shapes contribute zero findings (reported per-source in sources), never a crash.

MISSING / INSUFFICIENT DATA

• Zero recognisable findings → status="insufficient" with an empty roadmap and a reason (never a fabricated plan).
• No traffic anywhere → reach falls back to URL counts and has_traffic_data=false; scores are still comparable but confidence is lower. NEVER invent clicks or a severity the source did not provide.
• Unknown finding_type → default owner (by prefix), effort 3, impact 2, confidence without the "known type" bonus — surfaced honestly, not silently dropped.

OUTPUT

One JSON object per references/output.schema.json. No prose.

FILES

• scripts/remediation_prioritizer.py — shape detection/normalisation, dedupe, RICE/ICE scoring, phase sequencing, ticket payloads.
• references/output.schema.json — output contract.