decision-extraction

name	decision-extraction
description	Identify and formalize durable product and technical decisions from agent outputs into long-term memory. Activate after Discovery produces artefacts, Delivery resolves trade-offs, or product direction materially changes.
license	ELv2
compatibility	Works with any filesystem-based AI coding agent
metadata	{"author":"gaai-framework","version":"1.0","category":"cross","track":"cross-cutting","id":"SKILL-DECISION-EXTRACTION-001","updated_at":"2026-04-05T00:00:00.000Z","status":"stable"}
inputs	[{"recent_agent_outputs":"session outputs from the invoking agent, or file paths to artefacts produced in the current session (e.g., evaluation reports, refined stories, approach-evaluation outputs)"},"contexts/artefacts/** (governed)"]
outputs	["contexts/memory/decisions/DEC-{N}.md (individual ADR file)","contexts/memory/decisions/_log.md (next ID updated)","contexts/memory/index.md (registry + file count updated)"]

Decision Extraction

Purpose / When to Activate

Activate after:

• Discovery produces epics, scope clarifications, or priorities
• Delivery resolves technical trade-offs or architectural constraints
• QA surfaces systemic issues requiring policy decisions
• Product direction materially changes

Do NOT use for trivial steps, implementation details, brainstorming, or reversible micro-choices.

---

Process

0. Decision Consistency Gate (mandatory). Before extracting any new decision:

   • Read contexts/memory/index.md → scan the Decision Registry by domain to identify relevant existing decisions
   • Load the specific decisions/DEC-{ID}.md files for decisions in the affected domain(s)
   • Verify the proposed decision does NOT contradict any active decision
   • If contradiction found: either explicitly supersede (set superseded_by in old file + supersedes in new file) with rationale, or STOP and escalate to human. When escalating due to contradiction: before surfacing the escalation, grep contexts/ for all occurrences of DEC-{id} (where {id} is the contradicted decision's ID). Collect every file path that references the contradicted DEC — memory files, stories, and architecture docs. Present this impact list alongside the escalation message so the human can assess scope before deciding. If the contexts/ directory scan fails: proceed with the escalation without the impact list — the escalation is more important than the impact details.
   • If unable to determine consistency → STOP and escalate to human
   • Never record a decision silently if it may conflict with an existing one

1. Scan outputs for explicit or implicit decisions: architectural choices, accepted trade-offs, scope boundaries, prioritization shifts, constraints introduced

2. Filter strictly for durable, governance-relevant decisions

3. Deduplication check: Scan the Decision Registry in index.md for existing entries covering the same topic. If found: (a) if the new decision supersedes the old, update the old DEC-{ID}.md file's frontmatter (status: superseded, superseded_by: DEC-{new-id}) and record the supersession in the new entry's supersedes field; (b) if the new decision confirms the old, skip writing a duplicate. 3b. Cross-reference assignment: For the new decision, populate related_to with up to 5 DEC IDs that are directly related (same domain cluster, supersession chain, or shared concern). Only include decisions the new entry explicitly builds on, refines, or constrains. If no strong relation exists, leave as [].

4. Convert each into a structured ADR file (see Output Format below):

   • Context
   • Decision
   • Impact

5. Classify using the 10 canonical domains: architecture, matching, expert-system, billing, booking, infrastructure, strategy, governance, market, content. And 3 levels: strategic (WHAT/WHY), architectural (HOW), operational (PROCESS).

6. Get next available ID from decisions/_log.md → write decisions/DEC-{N}.md

7. Update _log.md: increment next available ID, add one-line entry for the new decision

8. MANDATORY GATE — Update Decision Registry (index.md or sibling index-decisions.md if extracted): Add one row per new decision. Row form is pointer-only (per memory-index-compact skill, SKILL-MEMORY-INDEX-COMPACT-001): | DEC-{N} | {domain} | {level} | {≤30-word topic + key relations e.g. amends/supersedes} | {status} {YYYY-MM-DD} |. HARD CAP : ≤ 200 chars per row. Forbidden in row : Tier 2 cycle trails (F-counts, REFINE narratives), commit SHAs, §-numbered substance dumps duplicating DEC body, validation ceremony prose, drift-heal forensic prose — these live in _log.md and the DEC body, NOT in the registry. Increment file count in Shared Categories table. Verify: re-read the row + count its chars ; if > 200 chars, compact before completing. Blocking gate — do not output success until confirmed.

9. Summary range check: Read the Summaries section of index.md. If the new DEC ID exceeds the highest DEC covered by the latest summary file (e.g., summary covers 90–155 but new DEC is 156), append a line to decisions/_log.md: # ⚠️ PENDING: extend summary range to DEC-{new-max-id} — run memory-refresh. This signals the next memory-refresh cycle to extend the summary. Do NOT create a new summary mid-delivery.

---

Output Format

Each decision is an individual ADR file: decisions/DEC-{N}.md (sequential numeric ID).

---
id: DEC-{N}
domain: architecture | matching | expert-system | billing | booking | infrastructure | strategy | governance | market | content
level: strategic | architectural | operational
title: "Decision Title"
status: active
created_by: discovery
created_at: YYYY-MM-DD
last_updated_by: discovery
last_updated_at: YYYY-MM-DD
supersedes: null          # or DEC-{old-id} if replacing
superseded_by: null
tags:
  - {relevant tags}
related_to: []            # optional — max 5 DEC IDs
---

# DEC-{N} — Decision Title

## Context
...

## Decision
...

## Impact
...

---

Quality Checks

• All major decisions become explicit memory
• No repeated reasoning across sessions
• Governance trail is traceable
• Memory grows only with high-signal knowledge

---

Non-Goals

This skill must NOT:

• Summarize entire sessions
• Capture raw logs
• Duplicate existing decisions
• Store trivial steps
• Invent interpretation without artefact support

If future agents benefit from knowing it → extract it. If not → do not store it. Memory is leverage — not history.