assumptions

SkillDatabases & data

Manage the assumption registry — track, verify, and query assumptions about external dependencies and system behavior. Prevents costly rediscovery of known failures.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the assumptions skill

What this skill tells your AI

The instructions your AI receives, as published by kastalien-research/thoughtbox in .agents/skills/assumptions/SKILL.md and read by ahel’s review.

Manage assumptions: $ARGUMENTS

Commands

Parse the first word of $ARGUMENTS to determine the command:

list — Show all tracked assumptions

  1. Read all .assumptions/*.jsonl files
  2. Parse each line as a JSON record
  3. Display sorted by confidence (lowest first) or staleness (oldest verification first)
  4. Format as a table: ID | Category | Claim | Confidence | Last Verified | Status

add — Register a new assumption

Parse remaining arguments for: --category, --claim, --evidence, --source

Create a new assumption record in .assumptions/registry.jsonl:

{
  "id": "<category>-<short-slug>",
  "category": "<api|dependency|behavior|environment|tooling>",
  "claim": "<what we assume to be true>",
  "evidence": "<what supports this assumption>",
  "source": "<where this was discovered>",
  "confidence": 0.8,
  "created": "<ISO 8601>",
  "last_verified": "<ISO 8601>",
  "verification_method": "<how to test this>",
  "failure_history": [],
  "status": "active",
  "blast_radius": "<what breaks if this assumption is wrong>"
}

verify — Re-verify an assumption

  1. Read the assumption record by ID
  2. Execute the verification method (may involve web search, API calls, or code checks)
  3. Update last_verified timestamp and confidence score
  4. If verification fails, add to failure_history and reduce confidence
  5. If confidence drops below 0.3, mark status as suspect and warn

stale — Show assumptions that need re-verification

  1. Read all assumption records
  2. Filter to those where last_verified is more than 14 days ago
  3. Sort by blast_radius (highest first)
  4. Display with suggested verification actions

seed — Seed registry from MEMORY.md gotchas

  1. Read MEMORY.md
  2. Extract entries from "Gotchas", "Known Bugs", and "MCP Knowledge API Gotchas" sections
  3. For each entry, create an assumption record with:
    • category: inferred from content (api, dependency, behavior, etc.)
    • claim: the gotcha statement
    • evidence: "Discovered empirically" + date from MEMORY.md
    • confidence: 0.9 (verified by experience)
    • verification_method: suggested test

Schema

Each assumption record:

FieldTypeRequiredDescription
idstringyesUnique identifier (category-slug format)
categoryenumyesapi, dependency, behavior, environment, tooling
claimstringyesWhat we assume to be true
evidencestringyesWhat supports this claim
sourcestringyesWhere this was discovered (session, test, docs)
confidencefloatyes0.0 to 1.0 confidence score
createdstringyesISO 8601 creation timestamp
last_verifiedstringyesISO 8601 last verification timestamp
verification_methodstringnoHow to re-test this assumption
failure_historyarraynoPast verification failures with timestamps and details
statusenumyesactive, suspect, retired, verified
blast_radiusstringnoWhat breaks if this assumption is wrong
dependenciesarraynoOther assumptions this depends on

Output

Always end with a summary:

## Assumption Registry Status

Total: {N} assumptions
Active: {N} | Suspect: {N} | Retired: {N}
Stale (>14 days): {N}
Highest blast radius unverified: {assumption_id}

Signals

GitHub stars
64
Forks
20
Last commit
Jul 2026
Advanced
Catalog kind
skill
Gateway key
assumptions
Source
github.com/kastalien-research/thoughtbox