Observation Triage Cycle

This skill defines the SCAN-TRIAGE-PROMOTE cycle for processing archived observations and routing them to actionable outcomes.

When to Use

Manual invocation: Skill(skill="observation-triage-cycle") when processing accumulated observations. Via recipe: just triage-observations for quick scan. Scheduled: Can be invoked during heartbeat or session-end.

The Cycle

SCAN --> TRIAGE --> PROMOTE
  |         |           |
  |         |           +-> spawn:INV, spawn:WORK, spawn:FIX
  |         |           +-> memory (store insight)
  |         |           +-> discuss (flag for operator)
  |         |           +-> dismiss (no action)
  |         |
  |         +-> Apply dimensions: category, action, priority
  |
  +-> Find archived observations with triage_status: pending

1. SCAN Phase

Goal: Find observations from both memory (retro-cycle outputs) and filesystem (legacy observations.md).

Actions:

1a. Memory Scan (retro-triage, WORK-143)

Query memory for retro-cycle outputs using MCP db_query tool:

K/S/S Directives:

mcp__haios-memory__db_query(sql="SELECT id, content FROM concepts WHERE type = 'Directive' AND (content LIKE 'KEEP-%' OR content LIKE 'STOP-%' OR content LIKE 'START-%') ORDER BY id")

Bug Candidates:

mcp__haios-memory__db_query(sql="SELECT id, content FROM concepts WHERE type = 'Critique' AND content LIKE 'BUG %' ORDER BY id DESC")

Feature Candidates:

mcp__haios-memory__db_query(sql="SELECT id, content FROM concepts WHERE type = 'Critique' AND content LIKE 'FEATURE %' ORDER BY id DESC")

Refactor Candidates:

mcp__haios-memory__db_query(sql="SELECT id, content FROM concepts WHERE type = 'Critique' AND content LIKE 'REFACTOR %' ORDER BY id DESC")

Upgrade Candidates:

mcp__haios-memory__db_query(sql="SELECT id, content FROM concepts WHERE type = 'Critique' AND content LIKE 'UPGRADE %' ORDER BY id DESC")

Report counts for each category.

1b. Filesystem Scan (legacy)

just triage-observations

Report count of items with pending observations.

Exit Criteria:

Memory scan completed (K/S/S, bugs, features, refactors, upgrades)
Filesystem scan completed
Counts reported for both sources

Tools: mcp__haios-memory__db_query, Bash(just triage-observations), Read

2. TRIAGE Phase

Goal: Process observations from both memory and filesystem sources.

2a. Retro-Triage (memory source, WORK-143)

K/S/S Frequency Aggregation: From K/S/S directive query results, group by directive text (normalized):

Count frequency of each Keep/Stop/Start directive across entries
Format: KEEP-N: prefix → category=keep, directive=text after prefix
Rank by frequency descending — high-frequency directives are strongest signals
Present aggregated K/S/S to operator with counts

Bug Candidate Surfacing: From bug query results, parse BUG (confidence): description format:

Extract confidence tag (high/medium/low) and description
Sort by confidence (high first)
Present to operator for disposition (spawn:WORK, spawn:FIX, dismiss)

Feature Candidate Surfacing: From feature query results, parse FEATURE (confidence): description format:

Extract confidence tag and description
Sort by confidence (high first)
Present to operator for epoch-level triage (spawn:WORK, discuss, dismiss)

Refactor Candidate Surfacing: From refactor query results, parse REFACTOR (confidence): description format:

Extract confidence tag and description
Sort by confidence (high first)
Refactors are typically bundled with related work — present to operator for bundling decision (spawn:WORK, bundle-with:{id}, discuss, dismiss)

Upgrade Candidate Surfacing: From upgrade query results, parse UPGRADE (confidence): description format:

Extract confidence tag and description
Sort by confidence (high first)
Upgrades are prioritised against capacity — present to operator for capacity triage (spawn:WORK, discuss, dismiss)

Suggested Priority Consumption: Retro-extract items now include suggested_priority: now|next|later with rationale. When surfacing candidates, display the suggested priority alongside confidence to give the operator a head start. Triage makes the final call — the suggestion is context from the agent that did the work, not a binding decision.

2b. Legacy Triage (filesystem source)

Triage Dimensions (Industry Standard):

Dimension	Question	Values
Category	What type?	bug, gap, debt, insight, question, noise
Action	What to do?	spawn:INV, spawn:WORK, spawn:FIX, memory, discuss, dismiss
Priority	When?	P0 (blocking), P1 (this session), P2 (next), P3 (backlog)

For each filesystem observation, prompt for dimensions:

Category: What type of observation is this?
Action: What should happen next?
Priority: How urgent is this?

Example Triage Prompt:

Observation: "commit-close recipe uses legacy patterns"
Section: Gaps Noticed
Source: E2-217

Category? [bug/gap/debt/insight/question/noise]: gap
Action? [spawn:INV/spawn:WORK/spawn:FIX/memory/discuss/dismiss]: spawn:FIX
Priority? [P0/P1/P2/P3]: P2

Exit Criteria:

Retro-triage: K/S/S aggregation presented, bug/feature/refactor/upgrade candidates surfaced
Legacy triage: All filesystem observations triaged with valid dimensions
Combined triaged list ready for promotion

Tools: AskUserQuestion (for dimension selection and disposition), triage_observation()

3. PROMOTE Phase

Goal: Execute the action for each triaged observation.

Action Handlers:

Action	Execution
`spawn:INV`	Create investigation via `/new-investigation`
`spawn:WORK`	Create work item via `/new-work`
`spawn:FIX`	Create quick fix work item (priority P1)
`memory`	Store insight via `ingester_ingest`
`discuss`	Flag for operator (report but no action)
`dismiss`	Mark as noise (no action needed)

Actions:

For each triaged observation, execute action
Track results (spawned IDs, memory IDs)
Update observations.md with triage log entry
Set triage_status: triaged in frontmatter

Exit Criteria:

All actions executed
Observations.md updated with triage log
triage_status updated to triaged

Tools: /new-investigation, /new-work, ingester_ingest, Edit

Composition Map

Phase	Primary Tool	Output
SCAN (memory)	mcp__haios-memory__db_query	K/S/S directives, bug/feature/refactor/upgrade candidates
SCAN (filesystem)	Bash, Read	List of pending observations
TRIAGE (retro)	AskUserQuestion	KSS aggregation, candidate disposition
TRIAGE (legacy)	AskUserQuestion	Triaged observation list
PROMOTE	/new-*, ingester	Action results

Quick Reference

Phase	Question to Ask	If NO
SCAN	Any pending observations?	Exit - nothing to triage
TRIAGE	All dimensions valid?	Reprompt with validation error
PROMOTE	Action executed?	Report error, continue to next

Key Design Decisions

Decision	Choice	Rationale
Standalone skill	Not embedded in close-work-cycle	Keeps closure lean; triage runs independently
Interactive triage	Agent prompts per observation	Avoids automated misclassification
triage_status field	In frontmatter	Enables scan to skip processed files
Industry dimensions	bug/gap/debt/insight/question/noise	Aligns with bug triage patterns
Dual-source scan (WORK-143)	Memory + filesystem	Retro-cycle stores to memory; legacy observations in files. Both sources feed triage.
Content-pattern queries (WORK-143)	Query by type + content prefix, NOT source_adr	A1 critique: ingester_ingest source_path not stored in source_adr. Content patterns verified against actual data.
MCP db_query for memory scan (WORK-143)	Agent calls db_query directly	Python functions use db_query_fn injection for testability. Agent uses MCP tool at runtime.
4-type EXTRACT taxonomy	bug/feature/refactor/upgrade	Types route differently: bugs urgent, features backlog, refactors bundled, upgrades capacity-gated. Aligns with retro-cycle EXTRACT expansion.
Suggested priority consumption	Display retro-extract suggested_priority	Agent has freshest urgency context. Triage displays but overrides. Hybrid (Option C).

E2-217: Observation Capture Gate (capture mechanism)
INV-047: Close Cycle Observation Phase Ordering (capture timing)
INV-023: ReasoningBank Feedback Loop Architecture (same pattern)
close-work-cycle skill: Captures observations during closure
retro-cycle skill: Produces typed outputs stored to memory (WORK-142)
observations.py: Core functions for scan, parse, triage, promote, and retro-triage (WORK-143)
WORK-143: Retro-triage consumer update for memory-based provenance tags

observation-triage-cycle

Observation Triage Cycle

When to Use

The Cycle

1. SCAN Phase

1a. Memory Scan (retro-triage, WORK-143)

1b. Filesystem Scan (legacy)

2. TRIAGE Phase

2a. Retro-Triage (memory source, WORK-143)

2b. Legacy Triage (filesystem source)

3. PROMOTE Phase

Composition Map

Quick Reference

Key Design Decisions

Related