Compare commits
1 Commits
fix/11-boo
...
fix/14-mea
| Author | SHA1 | Date | |
|---|---|---|---|
|
bf003cd944 |
164
knowledge/SCHEMA.md
Normal file
164
knowledge/SCHEMA.md
Normal file
@@ -0,0 +1,164 @@
|
||||
# Knowledge File Format Specification
|
||||
|
||||
**Version:** 1
|
||||
**Issue:** #10
|
||||
**Status:** Draft
|
||||
|
||||
---
|
||||
|
||||
## Overview
|
||||
|
||||
The knowledge system has two layers:
|
||||
|
||||
1. **index.json** — Machine-readable fact index. Fast lookups by ID, category, repo, tags.
|
||||
2. **Knowledge files** (YAML) — Human-readable, editable facts organized by domain.
|
||||
|
||||
The harvester writes to both. The bootstrapper reads from index.json. Humans edit the YAML files directly.
|
||||
|
||||
---
|
||||
|
||||
## index.json Schema
|
||||
|
||||
```json
|
||||
{
|
||||
"version": 1,
|
||||
"last_updated": "ISO-8601 timestamp",
|
||||
"total_facts": 0,
|
||||
"facts": []
|
||||
}
|
||||
```
|
||||
|
||||
### Fact Object
|
||||
|
||||
| Field | Type | Required | Description |
|
||||
|-------|------|----------|-------------|
|
||||
| `id` | string | yes | Unique identifier: `{domain}:{category}:{sequence}` |
|
||||
| `fact` | string | yes | One-sentence description of the knowledge |
|
||||
| `category` | enum | yes | One of: `fact`, `pitfall`, `pattern`, `tool-quirk`, `question` |
|
||||
| `domain` | string | yes | Where this applies: repo name, `global`, or agent name |
|
||||
| `confidence` | float | yes | 0.0-1.0. How certain is this knowledge? |
|
||||
| `tags` | string[] | no | Searchable labels: `["git", "auth", "gitea"]` |
|
||||
| `source_count` | int | no | How many sessions confirmed this fact |
|
||||
| `first_seen` | date | no | ISO-8601 date first extracted |
|
||||
| `last_confirmed` | date | no | ISO-8601 date last seen in a session |
|
||||
| `expires` | date | no | Optional. After this date, fact is stale |
|
||||
| `related` | string[] | no | IDs of related facts |
|
||||
|
||||
### ID Format
|
||||
|
||||
```
|
||||
{domain}:{category}:{sequence}
|
||||
```
|
||||
|
||||
- `domain` — repo name, `global`, or agent type
|
||||
- `category` — one of the 5 categories
|
||||
- `sequence` — zero-padded 3-digit number: `001`, `002`, ...
|
||||
|
||||
Examples:
|
||||
- `the-nexus:pitfall:001`
|
||||
- `global:tool-quirk:012`
|
||||
- `hermes-agent:pattern:003`
|
||||
|
||||
### Categories
|
||||
|
||||
| Category | Definition | Example |
|
||||
|----------|------------|---------|
|
||||
| `fact` | Concrete, verifiable information | "Gitea API requires token auth at /api/v1" |
|
||||
| `pitfall` | Errors, wrong assumptions, time-wasters | "Assumed env var GITEA_TOKEN; actual path is ~/.config/gitea/token" |
|
||||
| `pattern` | Successful sequences of actions | "To deploy: test -> build -> push -> webhook" |
|
||||
| `tool-quirk` | Environment-specific behaviors | "URL format requires trailing slash on macOS" |
|
||||
| `question` | Identified but unanswered | "Need optimal batch size for harvesting" |
|
||||
|
||||
### Confidence Scoring
|
||||
|
||||
| Range | Meaning |
|
||||
|-------|---------|
|
||||
| 0.9-1.0 | Explicitly stated and verified |
|
||||
| 0.7-0.8 | Clearly implied by multiple data points |
|
||||
| 0.5-0.6 | Suggested but not fully verified |
|
||||
| 0.3-0.4 | Inferred from limited data |
|
||||
| 0.1-0.2 | Speculative or uncertain |
|
||||
|
||||
---
|
||||
|
||||
## Knowledge Files (YAML)
|
||||
|
||||
Human-readable files stored in `knowledge/` subdirectories.
|
||||
|
||||
### Directory Structure
|
||||
|
||||
```
|
||||
knowledge/
|
||||
├── index.json # Machine-readable fact index
|
||||
├── SCHEMA.md # This file
|
||||
├── global/ # Cross-repo knowledge
|
||||
│ ├── pitfalls.yaml # Pitfalls that span multiple repos
|
||||
│ ├── patterns.yaml # Proven workflows
|
||||
│ └── tool-quirks.yaml # Environment behaviors
|
||||
├── repos/ # Per-repo knowledge
|
||||
│ ├── the-nexus.yaml
|
||||
│ ├── hermes-agent.yaml
|
||||
│ └── ...
|
||||
└── agents/ # Agent-type knowledge
|
||||
├── mimo-sprint.yaml
|
||||
└── ...
|
||||
```
|
||||
|
||||
### YAML File Format
|
||||
|
||||
```yaml
|
||||
---
|
||||
domain: global # or repo name or agent name
|
||||
category: tool-quirk # fact, pitfall, pattern, tool-quirk, question
|
||||
version: 1
|
||||
last_updated: "2026-04-13"
|
||||
---
|
||||
|
||||
# Tool Quirks (Global)
|
||||
|
||||
Cross-environment behaviors that bite you if you don't know them.
|
||||
|
||||
## Authentication
|
||||
|
||||
- id: global:tool-quirk:001
|
||||
fact: "Gitea token stored at ~/.config/gitea/token, not env var"
|
||||
confidence: 0.95
|
||||
tags: [git, auth, gitea]
|
||||
source_count: 23
|
||||
first_seen: "2026-03-27"
|
||||
last_confirmed: "2026-04-13"
|
||||
related: [global:pitfall:003]
|
||||
```
|
||||
|
||||
### Rules
|
||||
|
||||
1. **One file per domain per category.** `repos/the-nexus.yaml` holds all the-nexus facts.
|
||||
2. **Markdown sections for humans.** YAML items live under markdown headers for Gitea UI readability.
|
||||
3. **ID is the link.** The `id` field connects YAML facts to index.json entries.
|
||||
4. **Harvester writes, humans edit.** Harvester appends. Humans correct confidence, add tags, mark expired.
|
||||
|
||||
---
|
||||
|
||||
## Sync Rules
|
||||
|
||||
1. **Harvester -> YAML:** Appends new facts to the appropriate YAML file.
|
||||
2. **Harvester -> index.json:** Adds/updates fact entries.
|
||||
3. **Human edits YAML:** Changes propagate to index.json on next harvester run.
|
||||
4. **Confidence decay:** Facts not confirmed in 30+ sessions get confidence *= 0.9.
|
||||
5. **Expiration:** Facts with `expires` date past current date are marked `stale`.
|
||||
|
||||
---
|
||||
|
||||
## Validation
|
||||
|
||||
Facts must pass these checks:
|
||||
|
||||
1. `id` matches format `{domain}:{category}:{sequence}`
|
||||
2. `category` is one of the 5 allowed values
|
||||
3. `confidence` is between 0.0 and 1.0
|
||||
4. `fact` is non-empty string, max 280 characters
|
||||
5. `domain` is non-empty string
|
||||
6. `tags` are lowercase alphanumeric + hyphens
|
||||
7. No duplicate IDs in index.json
|
||||
|
||||
Validation script: `scripts/validate_knowledge.py`
|
||||
80
knowledge/global/pitfalls.yaml
Normal file
80
knowledge/global/pitfalls.yaml
Normal file
@@ -0,0 +1,80 @@
|
||||
---
|
||||
domain: global
|
||||
category: pitfall
|
||||
version: 1
|
||||
last_updated: "2026-04-13"
|
||||
---
|
||||
|
||||
# Pitfalls (Global)
|
||||
|
||||
Cross-repo traps that waste time across the fleet.
|
||||
|
||||
## Git & Forge
|
||||
|
||||
- id: global:pitfall:001
|
||||
fact: "Branch protection requires 1 approval on main - API merges fail with 405 without it"
|
||||
confidence: 0.95
|
||||
tags: [git, merge, branch-protection, gitea]
|
||||
source_count: 12
|
||||
first_seen: "2026-04-05"
|
||||
last_confirmed: "2026-04-13"
|
||||
related: [the-nexus:pitfall:001]
|
||||
|
||||
- id: global:pitfall:002
|
||||
fact: "Never use --no-verify on git commits - it bypasses all hooks including safety checks"
|
||||
confidence: 0.95
|
||||
tags: [git, hooks, safety]
|
||||
source_count: 5
|
||||
first_seen: "2026-03-28"
|
||||
last_confirmed: "2026-04-13"
|
||||
|
||||
- id: global:pitfall:003
|
||||
fact: "Gitea PR creation workaround needed on the-nexus - direct API call fails, use alternative endpoint"
|
||||
confidence: 0.9
|
||||
tags: [gitea, pr, api, workaround]
|
||||
source_count: 4
|
||||
first_seen: "2026-04-06"
|
||||
last_confirmed: "2026-04-12"
|
||||
|
||||
## Agent Operations
|
||||
|
||||
- id: global:pitfall:004
|
||||
fact: "Anthropic is BANNED from fallback chain - if fallback triggers to Anthropic, something is wrong"
|
||||
confidence: 0.95
|
||||
tags: [provider, anthropic, fallback]
|
||||
source_count: 7
|
||||
first_seen: "2026-03-30"
|
||||
last_confirmed: "2026-04-13"
|
||||
|
||||
- id: global:pitfall:005
|
||||
fact: "Telegram tokens expired - don't assume Telegram notifications work without checking"
|
||||
confidence: 0.85
|
||||
tags: [telegram, notifications, token]
|
||||
source_count: 3
|
||||
first_seen: "2026-04-02"
|
||||
|
||||
- id: global:pitfall:006
|
||||
fact: "Multiple gateways = 'cannot schedule futures' error - only one gateway process should run"
|
||||
confidence: 0.9
|
||||
tags: [gateway, cron, process]
|
||||
source_count: 4
|
||||
first_seen: "2026-04-04"
|
||||
last_confirmed: "2026-04-11"
|
||||
|
||||
## Testing
|
||||
|
||||
- id: global:pitfall:007
|
||||
fact: "pytest root collection picks up operational *_test.py scripts - restrict to tests/ directory"
|
||||
confidence: 0.9
|
||||
tags: [pytest, test, collection]
|
||||
source_count: 3
|
||||
first_seen: "2026-04-07"
|
||||
last_confirmed: "2026-04-13"
|
||||
|
||||
- id: global:pitfall:008
|
||||
fact: "TDD: test 1 before building 55 - verify the cycle works before scaling"
|
||||
confidence: 0.95
|
||||
tags: [tdd, testing, methodology]
|
||||
source_count: 8
|
||||
first_seen: "2026-03-25"
|
||||
last_confirmed: "2026-04-13"
|
||||
73
knowledge/global/tool-quirks.yaml
Normal file
73
knowledge/global/tool-quirks.yaml
Normal file
@@ -0,0 +1,73 @@
|
||||
---
|
||||
domain: global
|
||||
category: tool-quirk
|
||||
version: 1
|
||||
last_updated: "2026-04-13"
|
||||
---
|
||||
|
||||
# Tool Quirks (Global)
|
||||
|
||||
Cross-environment behaviors that bite you if you don't know them.
|
||||
|
||||
## Authentication
|
||||
|
||||
- id: global:tool-quirk:001
|
||||
fact: "Gitea token stored at ~/.config/gitea/token, not env var GITEA_TOKEN"
|
||||
confidence: 0.95
|
||||
tags: [git, auth, gitea, token]
|
||||
source_count: 23
|
||||
first_seen: "2026-03-27"
|
||||
last_confirmed: "2026-04-13"
|
||||
related: [global:pitfall:001]
|
||||
|
||||
- id: global:tool-quirk:002
|
||||
fact: "Gitea API uses 'Authorization: token TOKEN' header format, not Bearer"
|
||||
confidence: 0.9
|
||||
tags: [git, api, gitea]
|
||||
source_count: 8
|
||||
first_seen: "2026-03-28"
|
||||
last_confirmed: "2026-04-12"
|
||||
|
||||
- id: global:tool-quirk:003
|
||||
fact: "Gitea Issues API type=issues param does NOT filter PRs - use truthiness check on pull_request field"
|
||||
confidence: 0.95
|
||||
tags: [gitea, api, issues, pr]
|
||||
source_count: 6
|
||||
first_seen: "2026-04-01"
|
||||
last_confirmed: "2026-04-13"
|
||||
|
||||
## Paths & Environment
|
||||
|
||||
- id: global:tool-quirk:004
|
||||
fact: "~/.hermes is the default hermes home - check get_hermes_home() not the path literal"
|
||||
confidence: 0.9
|
||||
tags: [paths, hermes, env]
|
||||
source_count: 10
|
||||
first_seen: "2026-03-30"
|
||||
last_confirmed: "2026-04-13"
|
||||
related: [hermes-agent:pitfall:005]
|
||||
|
||||
- id: global:tool-quirk:005
|
||||
fact: "Ansible vault-encrypted vars in YAML require vault_inline_vars plugin - standard ansible-vault fails"
|
||||
confidence: 0.85
|
||||
tags: [ansible, vault, config]
|
||||
source_count: 3
|
||||
first_seen: "2026-04-02"
|
||||
|
||||
## Model & Inference
|
||||
|
||||
- id: global:tool-quirk:006
|
||||
fact: "mimo-v2-pro via Nous Research is the default model - don't assume Anthropic is available"
|
||||
confidence: 0.95
|
||||
tags: [model, provider, nous, default]
|
||||
source_count: 15
|
||||
first_seen: "2026-03-25"
|
||||
last_confirmed: "2026-04-13"
|
||||
|
||||
- id: global:tool-quirk:007
|
||||
fact: "Kill + restart with 'hermes chat' preserves old model state - NEVER use --resume"
|
||||
confidence: 0.95
|
||||
tags: [hermes, model, restart, session]
|
||||
source_count: 8
|
||||
first_seen: "2026-03-29"
|
||||
last_confirmed: "2026-04-12"
|
||||
82
knowledge/repos/hermes-agent.yaml
Normal file
82
knowledge/repos/hermes-agent.yaml
Normal file
@@ -0,0 +1,82 @@
|
||||
---
|
||||
domain: hermes-agent
|
||||
category: pitfall
|
||||
version: 1
|
||||
last_updated: "2026-04-13"
|
||||
---
|
||||
|
||||
# Pitfalls (hermes-agent)
|
||||
|
||||
Things that go wrong in this repo if you don't know the traps.
|
||||
|
||||
## Cron & Deployment
|
||||
|
||||
- id: hermes-agent:pitfall:001
|
||||
fact: "deploy-crons.py leaves jobs in mixed model format - some have provider/model, some just model"
|
||||
confidence: 0.95
|
||||
tags: [cron, deploy, model, config]
|
||||
source_count: 5
|
||||
first_seen: "2026-04-08"
|
||||
last_confirmed: "2026-04-13"
|
||||
related: [hermes-agent:pitfall:002, hermes-agent:pitfall:003]
|
||||
|
||||
- id: hermes-agent:pitfall:002
|
||||
fact: "deploy-crons.py --deploy doesn't set legacy skill field from skills list, breaking older jobs"
|
||||
confidence: 0.9
|
||||
tags: [cron, deploy, skills]
|
||||
source_count: 3
|
||||
first_seen: "2026-04-09"
|
||||
last_confirmed: "2026-04-13"
|
||||
related: [hermes-agent:pitfall:001]
|
||||
|
||||
- id: hermes-agent:pitfall:003
|
||||
fact: "Cron jobs with blank fallback_model fields trigger spurious gateway warnings"
|
||||
confidence: 0.9
|
||||
tags: [cron, model, fallback]
|
||||
source_count: 4
|
||||
first_seen: "2026-04-07"
|
||||
last_confirmed: "2026-04-12"
|
||||
related: [hermes-agent:pitfall:001]
|
||||
|
||||
- id: hermes-agent:pitfall:004
|
||||
fact: "model-watchdog.py checks first provider line, not model.provider - causes false drift alarms"
|
||||
confidence: 0.9
|
||||
tags: [watchdog, model, config]
|
||||
source_count: 3
|
||||
first_seen: "2026-04-08"
|
||||
last_confirmed: "2026-04-13"
|
||||
|
||||
## Path & Environment
|
||||
|
||||
- id: hermes-agent:pitfall:005
|
||||
fact: "10+ files read HERMES_HOME directly instead of get_hermes_home() - breaks on custom paths"
|
||||
confidence: 0.85
|
||||
tags: [paths, env, hermes-home]
|
||||
source_count: 6
|
||||
first_seen: "2026-04-06"
|
||||
last_confirmed: "2026-04-12"
|
||||
related: [global:pitfall:002]
|
||||
|
||||
- id: hermes-agent:pitfall:006
|
||||
fact: "get_hermes_home() doesn't expand tilde when HERMES_HOME=~/... is set"
|
||||
confidence: 0.8
|
||||
tags: [paths, env, bug]
|
||||
source_count: 2
|
||||
first_seen: "2026-04-05"
|
||||
|
||||
## SSH & Dispatch
|
||||
|
||||
- id: hermes-agent:pitfall:007
|
||||
fact: "vps-agent-dispatch reports OK while remote hermes binary path is broken"
|
||||
confidence: 0.9
|
||||
tags: [ssh, dispatch, vps]
|
||||
source_count: 4
|
||||
first_seen: "2026-04-07"
|
||||
last_confirmed: "2026-04-11"
|
||||
|
||||
- id: hermes-agent:pitfall:008
|
||||
fact: "nightwatch-health-monitor SSH check fails on cloud-model-only deployments"
|
||||
confidence: 0.85
|
||||
tags: [ssh, health, cloud]
|
||||
source_count: 2
|
||||
first_seen: "2026-04-10"
|
||||
65
knowledge/repos/the-nexus.yaml
Normal file
65
knowledge/repos/the-nexus.yaml
Normal file
@@ -0,0 +1,65 @@
|
||||
---
|
||||
domain: the-nexus
|
||||
category: pitfall
|
||||
version: 1
|
||||
last_updated: "2026-04-13"
|
||||
---
|
||||
|
||||
# Pitfalls (the-nexus)
|
||||
|
||||
Things that go wrong in this repo if you don't know the traps.
|
||||
|
||||
## Git & Merging
|
||||
|
||||
- id: the-nexus:pitfall:001
|
||||
fact: "Merges fail with HTTP 405 due to branch protection - must use merge API with 1 approval"
|
||||
confidence: 0.95
|
||||
tags: [git, merge, branch-protection, gitea]
|
||||
source_count: 12
|
||||
first_seen: "2026-04-05"
|
||||
last_confirmed: "2026-04-13"
|
||||
related: [global:pitfall:001]
|
||||
|
||||
- id: the-nexus:pitfall:002
|
||||
fact: "ThreadingHTTPServer required for multi-user bridge - standard HTTPServer blocks on concurrent requests"
|
||||
confidence: 0.95
|
||||
tags: [server, concurrency, bridge]
|
||||
source_count: 5
|
||||
first_seen: "2026-04-10"
|
||||
last_confirmed: "2026-04-13"
|
||||
|
||||
- id: the-nexus:pitfall:003
|
||||
fact: "ChatLog.log() crashes on message persistence when index.html has orphaned button tags"
|
||||
confidence: 0.9
|
||||
tags: [html, crash, chatlog]
|
||||
source_count: 3
|
||||
first_seen: "2026-04-12"
|
||||
last_confirmed: "2026-04-13"
|
||||
|
||||
## Three.js & Performance
|
||||
|
||||
- id: the-nexus:pitfall:004
|
||||
fact: "Three.js LOD not implemented - local hardware struggles with full scene without texture optimization"
|
||||
confidence: 0.85
|
||||
tags: [threejs, performance, lod]
|
||||
source_count: 4
|
||||
first_seen: "2026-04-09"
|
||||
last_confirmed: "2026-04-13"
|
||||
|
||||
- id: the-nexus:pitfall:005
|
||||
fact: "Duplicate content blocks appear in index.html when PR merges conflict silently"
|
||||
confidence: 0.8
|
||||
tags: [html, merge-conflict, duplicate]
|
||||
source_count: 3
|
||||
first_seen: "2026-04-11"
|
||||
last_confirmed: "2026-04-13"
|
||||
|
||||
## Deployment
|
||||
|
||||
- id: the-nexus:pitfall:006
|
||||
fact: "Unified HTTP + WebSocket server required for proper URL deployment - separate servers break CORS"
|
||||
confidence: 0.9
|
||||
tags: [deploy, websocket, http, cors]
|
||||
source_count: 4
|
||||
first_seen: "2026-04-10"
|
||||
last_confirmed: "2026-04-13"
|
||||
61
metrics/dashboard.md
Normal file
61
metrics/dashboard.md
Normal file
@@ -0,0 +1,61 @@
|
||||
# Compounding Intelligence Metrics
|
||||
**Generated:** 2026-04-14T18:12:26.469085+00:00
|
||||
|
||||
## knowledge_velocity
|
||||
New facts extracted per day. Higher = compounding loop working.
|
||||
|
||||
**Value:** 1.61 | **7d trend:** N/A --- (unknown)
|
||||
|
||||
- total_facts: 29
|
||||
- period_days: 18
|
||||
- new_facts: 29
|
||||
|
||||
## knowledge_coverage
|
||||
Percentage of domains/repos with 10+ facts. Measures breadth.
|
||||
|
||||
**Value:** 0.333 | **7d trend:** N/A --- (unknown)
|
||||
|
||||
- covered_domains: 1
|
||||
- total_domains: 3
|
||||
|
||||
## hit_rate
|
||||
Percentage of sessions referencing bootstrapped knowledge.
|
||||
|
||||
**Value:** 0.676 | **7d trend:** N/A --- (unknown)
|
||||
|
||||
- hit_sessions: 8058
|
||||
- total_sessions: 11922
|
||||
|
||||
## error_recurrence
|
||||
Ratio of recurring errors. Lower = fleet learning from mistakes.
|
||||
|
||||
**Value:** 0.17 | **7d trend:** N/A --- (unknown)
|
||||
|
||||
- unique_errors: 53615
|
||||
- recurring_errors: 9093
|
||||
|
||||
## task_completion
|
||||
Percentage of sessions ending with successful completion.
|
||||
|
||||
**Value:** 0.452 | **7d trend:** N/A --- (unknown)
|
||||
|
||||
- normal_end_rate: 0.56
|
||||
- completed: 5385
|
||||
- total: 11922
|
||||
|
||||
## first_try_success
|
||||
Percentage of sessions completed without backtracking.
|
||||
|
||||
**Value:** 0.818 | **7d trend:** N/A --- (unknown)
|
||||
|
||||
- avg_tool_msg_ratio: 0.392
|
||||
- sampled: 5923
|
||||
|
||||
## knowledge_age
|
||||
Freshness of knowledge store. 1.0 = all fresh, 0.0 = all stale.
|
||||
|
||||
**Value:** 0.973 | **7d trend:** N/A --- (unknown)
|
||||
|
||||
- avg_age_days: 2.4
|
||||
- stale_facts: 0
|
||||
- total_facts: 29
|
||||
127
metrics/latest_snapshot.json
Normal file
127
metrics/latest_snapshot.json
Normal file
@@ -0,0 +1,127 @@
|
||||
{
|
||||
"generated_at": "2026-04-14T18:12:26.469085+00:00",
|
||||
"knowledge_velocity": {
|
||||
"value": 1.61,
|
||||
"total_facts": 29,
|
||||
"period_days": 18,
|
||||
"new_facts": 29
|
||||
},
|
||||
"knowledge_coverage": {
|
||||
"value": 0.333,
|
||||
"covered_domains": 1,
|
||||
"total_domains": 3,
|
||||
"domain_details": {
|
||||
"global": 15,
|
||||
"hermes-agent": 8,
|
||||
"the-nexus": 6
|
||||
}
|
||||
},
|
||||
"hit_rate": {
|
||||
"value": 0.676,
|
||||
"hit_sessions": 8058,
|
||||
"total_sessions": 11922
|
||||
},
|
||||
"error_recurrence": {
|
||||
"value": 0.17,
|
||||
"unique_errors": 53615,
|
||||
"recurring_errors": 9093,
|
||||
"top_errors": [
|
||||
{
|
||||
"error": "s, report the error details.",
|
||||
"sessions": 1185
|
||||
},
|
||||
{
|
||||
"error": "\": \"traceback (most recent call last):\\n file \\\"/private/var/folders/9k/v07xkpp",
|
||||
"sessions": 695
|
||||
},
|
||||
{
|
||||
"error": "\", \"output\": \"\\n--- stderr ---\\ntraceback (most recent call last):\\n file \\\"/pr",
|
||||
"sessions": 684
|
||||
},
|
||||
{
|
||||
"error": "ures \u2192 file an issue with the traceback and tag [bug]",
|
||||
"sessions": 320
|
||||
},
|
||||
{
|
||||
"error": "s you encounter \u2192 file an issue with reproduction steps",
|
||||
"sessions": 320
|
||||
},
|
||||
{
|
||||
"error": "s, file a [bug] issue first.",
|
||||
"sessions": 320
|
||||
},
|
||||
{
|
||||
"error": ", fix the code.",
|
||||
"sessions": 314
|
||||
},
|
||||
{
|
||||
"error": "fix it before doing anything else.",
|
||||
"sessions": 313
|
||||
},
|
||||
{
|
||||
"error": "ures \u2192 add a review comment explaining what's wrong",
|
||||
"sessions": 303
|
||||
},
|
||||
{
|
||||
"error": "ures \u2014 they're your roadmap for guardrails",
|
||||
"sessions": 303
|
||||
}
|
||||
]
|
||||
},
|
||||
"task_completion": {
|
||||
"value": 0.452,
|
||||
"normal_end_rate": 0.56,
|
||||
"completed": 5385,
|
||||
"total": 11922,
|
||||
"breakdown": {
|
||||
"cron_complete": 5354,
|
||||
"unknown": 5248,
|
||||
"compression": 1092,
|
||||
"cli_close": 197,
|
||||
"session_reset": 31
|
||||
}
|
||||
},
|
||||
"first_try_success": {
|
||||
"value": 0.818,
|
||||
"avg_tool_msg_ratio": 0.392,
|
||||
"sampled": 5923,
|
||||
"interpretation": "Higher value = fewer backtracks = better first-try success"
|
||||
},
|
||||
"knowledge_age": {
|
||||
"value": 0.973,
|
||||
"avg_age_days": 2.4,
|
||||
"stale_facts": 0,
|
||||
"total_facts": 29,
|
||||
"interpretation": "1.0 = all facts fresh. 0.0 = all facts 90+ days old"
|
||||
},
|
||||
"trend_7d": {
|
||||
"knowledge_velocity": {
|
||||
"delta": "N/A",
|
||||
"direction": "unknown"
|
||||
},
|
||||
"knowledge_coverage": {
|
||||
"delta": "N/A",
|
||||
"direction": "unknown"
|
||||
},
|
||||
"hit_rate": {
|
||||
"delta": "N/A",
|
||||
"direction": "unknown"
|
||||
},
|
||||
"error_recurrence": {
|
||||
"delta": "N/A",
|
||||
"direction": "unknown"
|
||||
},
|
||||
"task_completion": {
|
||||
"delta": "N/A",
|
||||
"direction": "unknown"
|
||||
},
|
||||
"first_try_success": {
|
||||
"delta": "N/A",
|
||||
"direction": "unknown"
|
||||
},
|
||||
"knowledge_age": {
|
||||
"delta": "N/A",
|
||||
"direction": "unknown"
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -1,359 +0,0 @@
|
||||
#!/usr/bin/env python3
|
||||
"""
|
||||
Bootstrapper — assemble pre-session context from knowledge store.
|
||||
|
||||
Reads the knowledge store and produces a compact context block (2k tokens max)
|
||||
that can be injected into a new session so it starts with situational awareness.
|
||||
|
||||
Usage:
|
||||
python3 bootstrapper.py --repo the-nexus --agent mimo-sprint
|
||||
python3 bootstrapper.py --repo timmy-home --global
|
||||
python3 bootstrapper.py --global
|
||||
python3 bootstrapper.py --repo the-nexus --max-tokens 1000
|
||||
"""
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import sys
|
||||
from pathlib import Path
|
||||
from typing import Optional
|
||||
|
||||
# Resolve knowledge root relative to this script's parent
|
||||
SCRIPT_DIR = Path(__file__).resolve().parent
|
||||
REPO_ROOT = SCRIPT_DIR.parent
|
||||
KNOWLEDGE_DIR = REPO_ROOT / "knowledge"
|
||||
INDEX_PATH = KNOWLEDGE_DIR / "index.json"
|
||||
|
||||
# Approximate token count: ~4 chars per token for English text
|
||||
CHARS_PER_TOKEN = 4
|
||||
|
||||
# Category sort priority (lower = shown first)
|
||||
CATEGORY_PRIORITY = {
|
||||
"pitfall": 0,
|
||||
"tool-quirk": 1,
|
||||
"pattern": 2,
|
||||
"fact": 3,
|
||||
"question": 4,
|
||||
}
|
||||
|
||||
|
||||
def load_index(index_path: Path = INDEX_PATH) -> dict:
|
||||
"""Load and validate the knowledge index."""
|
||||
if not index_path.exists():
|
||||
return {"version": 1, "total_facts": 0, "facts": []}
|
||||
|
||||
with open(index_path) as f:
|
||||
data = json.load(f)
|
||||
|
||||
if "facts" not in data:
|
||||
print(f"WARNING: index.json missing 'facts' key", file=sys.stderr)
|
||||
return {"version": 1, "total_facts": 0, "facts": []}
|
||||
|
||||
return data
|
||||
|
||||
|
||||
def filter_facts(
|
||||
facts: list[dict],
|
||||
repo: Optional[str] = None,
|
||||
agent: Optional[str] = None,
|
||||
include_global: bool = True,
|
||||
) -> list[dict]:
|
||||
"""Filter facts by repo, agent, and global scope."""
|
||||
filtered = []
|
||||
|
||||
for fact in facts:
|
||||
fact_repo = fact.get("repo", "global")
|
||||
fact_agent = fact.get("agent", "")
|
||||
|
||||
# Match by repo (regardless of agent)
|
||||
if repo and fact_repo == repo:
|
||||
filtered.append(fact)
|
||||
continue
|
||||
|
||||
# Match by exact agent type
|
||||
if agent and fact_agent == agent:
|
||||
filtered.append(fact)
|
||||
continue
|
||||
|
||||
# Include global facts without agent restriction (universal facts)
|
||||
if include_global and fact_repo == "global" and not fact_agent:
|
||||
filtered.append(fact)
|
||||
|
||||
return filtered
|
||||
|
||||
|
||||
def sort_facts(facts: list[dict]) -> list[dict]:
|
||||
"""
|
||||
Sort facts by: confidence (desc), then category priority, then fact text.
|
||||
Most reliable and most dangerous facts come first.
|
||||
"""
|
||||
|
||||
def sort_key(f):
|
||||
confidence = f.get("confidence", 0.5)
|
||||
category = f.get("category", "fact")
|
||||
cat_priority = CATEGORY_PRIORITY.get(category, 5)
|
||||
return (-confidence, cat_priority, f.get("fact", ""))
|
||||
|
||||
return sorted(facts, key=sort_key)
|
||||
|
||||
|
||||
def load_repo_knowledge(repo: str) -> Optional[str]:
|
||||
"""Load per-repo knowledge markdown if it exists."""
|
||||
repo_path = KNOWLEDGE_DIR / "repos" / f"{repo}.md"
|
||||
if repo_path.exists():
|
||||
return repo_path.read_text().strip()
|
||||
return None
|
||||
|
||||
|
||||
def load_agent_knowledge(agent: str) -> Optional[str]:
|
||||
"""Load per-agent knowledge markdown if it exists."""
|
||||
agent_path = KNOWLEDGE_DIR / "agents" / f"{agent}.md"
|
||||
if agent_path.exists():
|
||||
return agent_path.read_text().strip()
|
||||
return None
|
||||
|
||||
|
||||
def load_global_knowledge() -> list[str]:
|
||||
"""Load all global knowledge markdown files."""
|
||||
global_dir = KNOWLEDGE_DIR / "global"
|
||||
if not global_dir.exists():
|
||||
return []
|
||||
|
||||
chunks = []
|
||||
for md_file in sorted(global_dir.glob("*.md")):
|
||||
content = md_file.read_text().strip()
|
||||
if content:
|
||||
chunks.append(content)
|
||||
return chunks
|
||||
|
||||
|
||||
def render_facts_section(facts: list[dict], category: str, label: str) -> str:
|
||||
"""Render a section of facts for a single category."""
|
||||
cat_facts = [f for f in facts if f.get("category") == category]
|
||||
if not cat_facts:
|
||||
return ""
|
||||
|
||||
lines = [f"### {label}\n"]
|
||||
for f in cat_facts:
|
||||
conf = f.get("confidence", 0.5)
|
||||
fact_text = f.get("fact", "")
|
||||
repo_tag = f.get("repo", "")
|
||||
if repo_tag and repo_tag != "global":
|
||||
lines.append(f"- [{conf:.0%}] ({repo_tag}) {fact_text}")
|
||||
else:
|
||||
lines.append(f"- [{conf:.0%}] {fact_text}")
|
||||
|
||||
return "\n".join(lines) + "\n"
|
||||
|
||||
|
||||
def estimate_tokens(text: str) -> int:
|
||||
"""Rough token estimate."""
|
||||
return len(text) // CHARS_PER_TOKEN
|
||||
|
||||
|
||||
def truncate_to_tokens(text: str, max_tokens: int) -> str:
|
||||
"""Truncate text to approximately max_tokens, cutting at line boundaries."""
|
||||
max_chars = max_tokens * CHARS_PER_TOKEN
|
||||
if len(text) <= max_chars:
|
||||
return text
|
||||
|
||||
# Cut at last newline before the limit
|
||||
truncated = text[:max_chars]
|
||||
last_newline = truncated.rfind("\n")
|
||||
if last_newline > 0:
|
||||
truncated = truncated[:last_newline]
|
||||
|
||||
return truncated + "\n\n[... truncated to fit context window ...]"
|
||||
|
||||
|
||||
def build_bootstrap_context(
|
||||
repo: Optional[str] = None,
|
||||
agent: Optional[str] = None,
|
||||
include_global: bool = True,
|
||||
max_tokens: int = 2000,
|
||||
index_path: Path = INDEX_PATH,
|
||||
) -> str:
|
||||
"""
|
||||
Build the full bootstrap context block.
|
||||
|
||||
Returns a markdown string suitable for injection into a session prompt.
|
||||
"""
|
||||
index = load_index(index_path)
|
||||
facts = index.get("facts", [])
|
||||
|
||||
# Filter
|
||||
filtered = filter_facts(facts, repo=repo, agent=agent, include_global=include_global)
|
||||
|
||||
# Sort
|
||||
sorted_facts = sort_facts(filtered)
|
||||
|
||||
# Build sections
|
||||
sections = ["## What You Know (bootstrapped)\n"]
|
||||
|
||||
# Per-repo markdown knowledge
|
||||
if repo:
|
||||
repo_md = load_repo_knowledge(repo)
|
||||
if repo_md:
|
||||
sections.append(f"### Repo Notes: {repo}\n")
|
||||
sections.append(repo_md + "\n")
|
||||
|
||||
# Structured facts by category
|
||||
if sorted_facts:
|
||||
# Group by source
|
||||
repo_facts = [f for f in sorted_facts if f.get("repo") == repo] if repo else []
|
||||
global_facts = [f for f in sorted_facts if f.get("repo") == "global"]
|
||||
agent_facts = [f for f in sorted_facts if f.get("agent") == agent] if agent else []
|
||||
|
||||
if repo_facts:
|
||||
sections.append(f"### Repo: {repo}\n")
|
||||
for cat, label in [
|
||||
("pitfall", "PITFALLS"),
|
||||
("tool-quirk", "QUIRKS"),
|
||||
("pattern", "PATTERNS"),
|
||||
("fact", "FACTS"),
|
||||
("question", "OPEN QUESTIONS"),
|
||||
]:
|
||||
section = render_facts_section(repo_facts, cat, label)
|
||||
if section:
|
||||
sections.append(section)
|
||||
|
||||
if global_facts:
|
||||
sections.append("### Global\n")
|
||||
for cat, label in [
|
||||
("pitfall", "PITFALLS"),
|
||||
("tool-quirk", "QUIRKS"),
|
||||
("pattern", "PATTERNS"),
|
||||
("fact", "FACTS"),
|
||||
]:
|
||||
section = render_facts_section(global_facts, cat, label)
|
||||
if section:
|
||||
sections.append(section)
|
||||
|
||||
if agent_facts:
|
||||
sections.append(f"### Agent Notes ({agent})\n")
|
||||
for cat, label in [
|
||||
("pitfall", "PITFALLS"),
|
||||
("tool-quirk", "QUIRKS"),
|
||||
("pattern", "PATTERNS"),
|
||||
("fact", "FACTS"),
|
||||
]:
|
||||
section = render_facts_section(agent_facts, cat, label)
|
||||
if section:
|
||||
sections.append(section)
|
||||
|
||||
# Per-agent markdown knowledge
|
||||
if agent:
|
||||
agent_md = load_agent_knowledge(agent)
|
||||
if agent_md:
|
||||
sections.append(f"### Agent Profile: {agent}\n")
|
||||
sections.append(agent_md + "\n")
|
||||
|
||||
# Global markdown knowledge
|
||||
global_chunks = load_global_knowledge()
|
||||
if global_chunks:
|
||||
sections.append("### Global Notes\n")
|
||||
sections.extend(chunk + "\n" for chunk in global_chunks)
|
||||
|
||||
# If nothing was found
|
||||
if len(sections) == 1:
|
||||
sections.append("_No relevant knowledge found. Starting fresh._\n")
|
||||
if not facts:
|
||||
sections.append(
|
||||
"_Knowledge store is empty. Run the harvester to populate it._\n"
|
||||
)
|
||||
|
||||
# Join and truncate
|
||||
context = "\n".join(sections)
|
||||
context = truncate_to_tokens(context, max_tokens)
|
||||
|
||||
return context
|
||||
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser(
|
||||
description="Assemble pre-session context from knowledge store"
|
||||
)
|
||||
parser.add_argument(
|
||||
"--repo",
|
||||
type=str,
|
||||
default=None,
|
||||
help="Repository name to filter facts by",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--agent",
|
||||
type=str,
|
||||
default=None,
|
||||
help="Agent type to filter facts by (e.g., mimo-sprint, groq-fast)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--global",
|
||||
dest="include_global",
|
||||
action="store_true",
|
||||
default=True,
|
||||
help="Include global facts (default: true)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--no-global",
|
||||
dest="include_global",
|
||||
action="store_false",
|
||||
help="Exclude global facts",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--max-tokens",
|
||||
type=int,
|
||||
default=2000,
|
||||
help="Maximum token count for output (default: 2000)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--index",
|
||||
type=str,
|
||||
default=None,
|
||||
help="Path to index.json (default: knowledge/index.json)",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--json",
|
||||
dest="output_json",
|
||||
action="store_true",
|
||||
help="Output raw JSON instead of markdown",
|
||||
)
|
||||
|
||||
args = parser.parse_args()
|
||||
|
||||
index_path = Path(args.index) if args.index else INDEX_PATH
|
||||
|
||||
if args.output_json:
|
||||
# JSON mode: return the filtered, sorted facts
|
||||
index = load_index(index_path)
|
||||
facts = index.get("facts", [])
|
||||
filtered = filter_facts(
|
||||
facts,
|
||||
repo=args.repo,
|
||||
agent=args.agent,
|
||||
include_global=args.include_global,
|
||||
)
|
||||
sorted_facts = sort_facts(filtered)
|
||||
output = {
|
||||
"repo": args.repo,
|
||||
"agent": args.agent,
|
||||
"include_global": args.include_global,
|
||||
"total_indexed": len(facts),
|
||||
"matched": len(sorted_facts),
|
||||
"facts": sorted_facts,
|
||||
}
|
||||
print(json.dumps(output, indent=2))
|
||||
else:
|
||||
# Markdown mode: full bootstrap context
|
||||
context = build_bootstrap_context(
|
||||
repo=args.repo,
|
||||
agent=args.agent,
|
||||
include_global=args.include_global,
|
||||
max_tokens=args.max_tokens,
|
||||
index_path=index_path,
|
||||
)
|
||||
print(context)
|
||||
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
sys.exit(main())
|
||||
403
scripts/measurer.py
Normal file
403
scripts/measurer.py
Normal file
@@ -0,0 +1,403 @@
|
||||
#!/usr/bin/env python3
|
||||
"""
|
||||
Compounding Intelligence Metrics Engine.
|
||||
|
||||
Computes 7 metrics that prove whether the knowledge compounding loop is working:
|
||||
1. Knowledge velocity -- new facts per day
|
||||
2. Knowledge coverage -- % of domains with >10 facts
|
||||
3. Hit rate -- % of sessions referencing bootstrap knowledge
|
||||
4. Error recurrence -- same errors across sessions (should decrease)
|
||||
5. Task completion -- % of sessions ending successfully
|
||||
6. First-try success -- actions without backtracking
|
||||
7. Knowledge age -- staleness of facts
|
||||
|
||||
Usage:
|
||||
python3 measurer.py # All metrics, all time
|
||||
python3 measurer.py --since 2026-04-01 # Time range
|
||||
python3 measurer.py --repo the-nexus # Per-repo metrics
|
||||
python3 measurer.py --format json # JSON output (default)
|
||||
python3 measurer.py --format markdown # Human-readable
|
||||
python3 measurer.py --knowledge-dir ./knowledge # Custom knowledge path
|
||||
python3 measurer.py --db ~/.hermes/state.db # Custom DB path
|
||||
|
||||
Data sources:
|
||||
- knowledge/index.json -- fact index
|
||||
- knowledge/ -- YAML fact files for coverage
|
||||
- ~/.hermes/state.db -- session/message metadata
|
||||
"""
|
||||
|
||||
import argparse
|
||||
import json
|
||||
import os
|
||||
import re
|
||||
import sqlite3
|
||||
import sys
|
||||
from collections import Counter, defaultdict
|
||||
from datetime import datetime, timedelta, timezone
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
|
||||
# --- Defaults ---
|
||||
|
||||
DEFAULT_KNOWLEDGE_DIR = Path(__file__).parent.parent / "knowledge"
|
||||
DEFAULT_DB_PATH = Path.home() / ".hermes" / "state.db"
|
||||
SEVEN_DAYS = timedelta(days=7)
|
||||
|
||||
|
||||
# --- Knowledge Store ---
|
||||
|
||||
def load_facts(knowledge_dir):
|
||||
index_path = knowledge_dir / "index.json"
|
||||
if not index_path.exists():
|
||||
return []
|
||||
with open(index_path) as f:
|
||||
data = json.load(f)
|
||||
return data.get("facts", [])
|
||||
|
||||
|
||||
def count_yaml_facts(knowledge_dir):
|
||||
domain_counts = {}
|
||||
for subdir in ["repos", "global", "agents"]:
|
||||
dirpath = knowledge_dir / subdir
|
||||
if not dirpath.exists():
|
||||
continue
|
||||
for yaml_file in dirpath.glob("*.yaml"):
|
||||
count = 0
|
||||
try:
|
||||
content = yaml_file.read_text()
|
||||
count = len(re.findall(r"^\s*-\s*id:", content, re.MULTILINE))
|
||||
except Exception:
|
||||
pass
|
||||
domain = yaml_file.stem
|
||||
domain_counts[domain] = domain_counts.get(domain, 0) + count
|
||||
return domain_counts
|
||||
|
||||
|
||||
# --- Session Database ---
|
||||
|
||||
def open_db(db_path):
|
||||
if not db_path.exists():
|
||||
print(f"WARNING: Database not found at {db_path}", file=sys.stderr)
|
||||
return None
|
||||
conn = sqlite3.connect(str(db_path))
|
||||
conn.row_factory = sqlite3.Row
|
||||
return conn
|
||||
|
||||
|
||||
def query_sessions(conn, since=None, repo=None):
|
||||
if conn is None:
|
||||
return []
|
||||
query = "SELECT id, started_at, ended_at, end_reason, message_count, tool_call_count, model FROM sessions WHERE 1=1"
|
||||
params = []
|
||||
if since:
|
||||
since_ts = datetime.fromisoformat(since).replace(tzinfo=timezone.utc).timestamp()
|
||||
query += " AND started_at >= ?"
|
||||
params.append(since_ts)
|
||||
query += " ORDER BY started_at ASC"
|
||||
cur = conn.execute(query, params)
|
||||
return [dict(row) for row in cur.fetchall()]
|
||||
|
||||
|
||||
def query_messages(conn, session_ids=None, since_ts=None):
|
||||
if conn is None:
|
||||
return []
|
||||
query = "SELECT m.session_id, m.role, m.content, m.tool_name, m.timestamp FROM messages m WHERE 1=1"
|
||||
params = []
|
||||
if since_ts:
|
||||
query += " AND m.timestamp >= ?"
|
||||
params.append(since_ts)
|
||||
if session_ids:
|
||||
placeholders = ",".join("?" for _ in session_ids)
|
||||
query += f" AND m.session_id IN ({placeholders})"
|
||||
params.extend(session_ids)
|
||||
cur = conn.execute(query, params)
|
||||
return [dict(row) for row in cur.fetchall()]
|
||||
|
||||
|
||||
# --- Metric Computations ---
|
||||
|
||||
def compute_knowledge_velocity(facts, since=None):
|
||||
if not facts:
|
||||
return {"value": 0.0, "total_facts": 0, "period_days": 0, "new_facts": 0}
|
||||
dates = []
|
||||
for f in facts:
|
||||
d = f.get("first_seen") or f.get("created")
|
||||
if d:
|
||||
try:
|
||||
dt = datetime.fromisoformat(d.replace("Z", "+00:00"))
|
||||
if dt.tzinfo is None:
|
||||
dt = dt.replace(tzinfo=timezone.utc)
|
||||
dates.append(dt)
|
||||
except (ValueError, AttributeError):
|
||||
pass
|
||||
if not dates:
|
||||
return {"value": 0.0, "total_facts": len(facts), "period_days": 0, "new_facts": 0}
|
||||
if since:
|
||||
cutoff = datetime.fromisoformat(since).replace(tzinfo=timezone.utc)
|
||||
dates = [d for d in dates if d >= cutoff]
|
||||
if not dates:
|
||||
return {"value": 0.0, "total_facts": len(facts), "period_days": 0, "new_facts": 0}
|
||||
earliest = min(dates)
|
||||
latest = max(dates)
|
||||
period_days = max((latest - earliest).days, 1)
|
||||
return {"value": round(len(dates) / period_days, 2), "total_facts": len(facts), "period_days": period_days, "new_facts": len(dates)}
|
||||
|
||||
|
||||
def compute_knowledge_coverage(facts, yaml_counts):
|
||||
domain_fact_counts = defaultdict(int)
|
||||
for f in facts:
|
||||
domain = f.get("domain", "unknown")
|
||||
domain_fact_counts[domain] += 1
|
||||
for domain, count in yaml_counts.items():
|
||||
domain_fact_counts[domain] = max(domain_fact_counts[domain], count)
|
||||
total_domains = len(domain_fact_counts)
|
||||
if total_domains == 0:
|
||||
return {"value": 0.0, "covered_domains": 0, "total_domains": 0, "domain_details": {}}
|
||||
covered = sum(1 for c in domain_fact_counts.values() if c >= 10)
|
||||
return {"value": round(covered / total_domains, 3), "covered_domains": covered, "total_domains": total_domains, "domain_details": dict(sorted(domain_fact_counts.items(), key=lambda x: -x[1])[:20])}
|
||||
|
||||
|
||||
def compute_hit_rate(sessions, messages, facts):
|
||||
if not sessions or not facts:
|
||||
return {"value": 0.0, "hit_sessions": 0, "total_sessions": len(sessions)}
|
||||
fact_fragments = set()
|
||||
for f in facts:
|
||||
text = f.get("fact", "").lower().strip()
|
||||
if len(text) > 10:
|
||||
fact_fragments.add(text)
|
||||
words = re.findall(r'\w{4,}', text)
|
||||
for w in words:
|
||||
fact_fragments.add(w)
|
||||
if not fact_fragments:
|
||||
return {"value": 0.0, "hit_sessions": 0, "total_sessions": len(sessions)}
|
||||
session_messages = defaultdict(list)
|
||||
for m in messages:
|
||||
content = (m.get("content") or "").lower()
|
||||
if content:
|
||||
session_messages[m["session_id"]].append(content)
|
||||
hit_sessions = 0
|
||||
for session in sessions:
|
||||
sid = session["id"]
|
||||
all_content = " ".join(session_messages.get(sid, []))
|
||||
if any(frag in all_content for frag in fact_fragments):
|
||||
hit_sessions += 1
|
||||
return {"value": round(hit_sessions / len(sessions), 3) if sessions else 0.0, "hit_sessions": hit_sessions, "total_sessions": len(sessions)}
|
||||
|
||||
|
||||
def compute_error_recurrence(messages):
|
||||
if not messages:
|
||||
return {"value": 0.0, "unique_errors": 0, "recurring_errors": 0, "top_errors": []}
|
||||
error_pattern = re.compile(r'(?:error|Error|ERROR|failed|FAIL|exception|Exception)[:\s]*(.{10,80})', re.IGNORECASE)
|
||||
error_to_sessions = defaultdict(set)
|
||||
for m in messages:
|
||||
content = m.get("content") or ""
|
||||
if not content:
|
||||
continue
|
||||
for match in error_pattern.finditer(content):
|
||||
sig = match.group(1).strip().lower()
|
||||
sig = re.sub(r'\s+', ' ', sig)
|
||||
if len(sig) > 5:
|
||||
error_to_sessions[sig].add(m["session_id"])
|
||||
if not error_to_sessions:
|
||||
return {"value": 0.0, "unique_errors": 0, "recurring_errors": 0, "top_errors": []}
|
||||
recurring = {e: s for e, s in error_to_sessions.items() if len(s) > 1}
|
||||
total_errors = len(error_to_sessions)
|
||||
recurring_count = len(recurring)
|
||||
top = sorted(recurring.items(), key=lambda x: -len(x[1]))[:10]
|
||||
return {"value": round(recurring_count / total_errors, 3) if total_errors else 0.0, "unique_errors": total_errors, "recurring_errors": recurring_count, "top_errors": [{"error": e, "sessions": len(s)} for e, s in top]}
|
||||
|
||||
|
||||
def compute_task_completion(sessions):
|
||||
if not sessions:
|
||||
return {"value": 0.0, "completed": 0, "total": 0, "breakdown": {}}
|
||||
breakdown = Counter()
|
||||
for s in sessions:
|
||||
reason = s.get("end_reason") or "unknown"
|
||||
breakdown[reason] += 1
|
||||
completed = breakdown.get("cron_complete", 0) + breakdown.get("session_reset", 0)
|
||||
normal_endings = completed + breakdown.get("cli_close", 0) + breakdown.get("compression", 0)
|
||||
return {"value": round(completed / len(sessions), 3) if sessions else 0.0, "normal_end_rate": round(normal_endings / len(sessions), 3) if sessions else 0.0, "completed": completed, "total": len(sessions), "breakdown": dict(breakdown.most_common())}
|
||||
|
||||
|
||||
def compute_first_try_success(sessions):
|
||||
if not sessions:
|
||||
return {"value": 0.0, "avg_tool_msg_ratio": 0.0, "sampled": 0}
|
||||
ratios = []
|
||||
for s in sessions:
|
||||
msgs = s.get("message_count", 0) or 0
|
||||
tools = s.get("tool_call_count", 0) or 0
|
||||
if msgs > 2:
|
||||
ratios.append(tools / msgs if msgs > 0 else 0)
|
||||
if not ratios:
|
||||
return {"value": 0.0, "avg_tool_msg_ratio": 0.0, "sampled": 0}
|
||||
avg_ratio = sum(ratios) / len(ratios)
|
||||
first_try = sum(1 for r in ratios if r < 0.5)
|
||||
return {"value": round(first_try / len(ratios), 3), "avg_tool_msg_ratio": round(avg_ratio, 3), "sampled": len(ratios), "interpretation": "Higher value = fewer backtracks = better first-try success"}
|
||||
|
||||
|
||||
def compute_knowledge_age(facts):
|
||||
if not facts:
|
||||
return {"value": 0.0, "avg_age_days": 0, "stale_facts": 0, "total_facts": 0}
|
||||
now = datetime.now(timezone.utc)
|
||||
ages = []
|
||||
stale_count = 0
|
||||
for f in facts:
|
||||
confirmed = f.get("last_confirmed") or f.get("first_seen")
|
||||
if confirmed:
|
||||
try:
|
||||
dt = datetime.fromisoformat(confirmed.replace("Z", "+00:00"))
|
||||
if dt.tzinfo is None:
|
||||
dt = dt.replace(tzinfo=timezone.utc)
|
||||
age = (now - dt).days
|
||||
ages.append(age)
|
||||
if age > 30:
|
||||
stale_count += 1
|
||||
except (ValueError, AttributeError):
|
||||
pass
|
||||
if not ages:
|
||||
return {"value": 0.0, "avg_age_days": 0, "stale_facts": 0, "total_facts": len(facts)}
|
||||
avg_age = sum(ages) / len(ages)
|
||||
freshness = max(0.0, 1.0 - (avg_age / 90))
|
||||
return {"value": round(freshness, 3), "avg_age_days": round(avg_age, 1), "stale_facts": stale_count, "total_facts": len(facts), "interpretation": "1.0 = all facts fresh. 0.0 = all facts 90+ days old"}
|
||||
|
||||
|
||||
# --- Trend Computation ---
|
||||
|
||||
def compute_trend(current, previous, metric_key="value"):
|
||||
if not previous:
|
||||
return {"delta": "N/A", "direction": "unknown"}
|
||||
curr_val = current.get(metric_key, 0)
|
||||
prev_val = previous.get(metric_key, 0)
|
||||
if prev_val == 0:
|
||||
return {"delta": "N/A (no baseline)", "direction": "unknown"}
|
||||
pct = ((curr_val - prev_val) / abs(prev_val)) * 100
|
||||
direction = "up" if pct > 0 else "down" if pct < 0 else "flat"
|
||||
if metric_key == "error_recurrence" or metric_key == "knowledge_age":
|
||||
direction_label = "good" if pct < 0 else "bad" if pct > 0 else "neutral"
|
||||
else:
|
||||
direction_label = "good" if pct > 0 else "bad" if pct < 0 else "neutral"
|
||||
return {"delta": f"{'+' if pct > 0 else ''}{pct:.1f}%", "direction": direction, "assessment": direction_label}
|
||||
|
||||
|
||||
# --- Output Formatters ---
|
||||
|
||||
def format_json(metrics):
|
||||
return json.dumps(metrics, indent=2)
|
||||
|
||||
|
||||
def format_markdown(metrics):
|
||||
lines = ["# Compounding Intelligence Metrics", f"**Generated:** {metrics.get('generated_at', 'unknown')}", ""]
|
||||
trend = metrics.get("trend_7d", {})
|
||||
|
||||
def metric_block(name, data, desc, good_direction="up"):
|
||||
val = data.get("value", 0)
|
||||
t = trend.get(name, {})
|
||||
delta = t.get("delta", "N/A")
|
||||
assessment = t.get("assessment", "unknown")
|
||||
arrow = "up" if assessment == "good" else "down" if assessment == "bad" else "---"
|
||||
lines.extend([f"## {name}", desc, "", f"**Value:** {val} | **7d trend:** {delta} {arrow} ({assessment})", ""])
|
||||
for k, v in data.items():
|
||||
if k != "value" and k != "interpretation":
|
||||
if isinstance(v, (int, float, str)):
|
||||
lines.append(f"- {k}: {v}")
|
||||
lines.append("")
|
||||
|
||||
metric_block("knowledge_velocity", metrics.get("knowledge_velocity", {}), "New facts extracted per day. Higher = compounding loop working.")
|
||||
metric_block("knowledge_coverage", metrics.get("knowledge_coverage", {}), "Percentage of domains/repos with 10+ facts. Measures breadth.")
|
||||
metric_block("hit_rate", metrics.get("hit_rate", {}), "Percentage of sessions referencing bootstrapped knowledge.")
|
||||
metric_block("error_recurrence", metrics.get("error_recurrence", {}), "Ratio of recurring errors. Lower = fleet learning from mistakes.", good_direction="down")
|
||||
metric_block("task_completion", metrics.get("task_completion", {}), "Percentage of sessions ending with successful completion.")
|
||||
metric_block("first_try_success", metrics.get("first_try_success", {}), "Percentage of sessions completed without backtracking.")
|
||||
metric_block("knowledge_age", metrics.get("knowledge_age", {}), "Freshness of knowledge store. 1.0 = all fresh, 0.0 = all stale.", good_direction="up")
|
||||
return "\n".join(lines)
|
||||
|
||||
|
||||
# --- Snapshot Persistence ---
|
||||
|
||||
def load_snapshot(metrics_dir):
|
||||
snapshot_path = metrics_dir / "latest_snapshot.json"
|
||||
if snapshot_path.exists():
|
||||
with open(snapshot_path) as f:
|
||||
return json.load(f)
|
||||
return {}
|
||||
|
||||
|
||||
def save_snapshot(metrics_dir, metrics):
|
||||
metrics_dir.mkdir(parents=True, exist_ok=True)
|
||||
snapshot_path = metrics_dir / "latest_snapshot.json"
|
||||
with open(snapshot_path, "w") as f:
|
||||
json.dump(metrics, f, indent=2)
|
||||
|
||||
|
||||
# --- Main ---
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser(description="Compounding Intelligence Metrics")
|
||||
parser.add_argument("--since", help="Start date (YYYY-MM-DD)")
|
||||
parser.add_argument("--repo", help="Filter by repo/domain")
|
||||
parser.add_argument("--format", choices=["json", "markdown"], default="json")
|
||||
parser.add_argument("--knowledge-dir", type=Path, default=DEFAULT_KNOWLEDGE_DIR)
|
||||
parser.add_argument("--db", type=Path, default=DEFAULT_DB_PATH)
|
||||
parser.add_argument("--save-snapshot", action="store_true", help="Save current metrics as snapshot for trend tracking")
|
||||
parser.add_argument("--metrics-dir", type=Path, default=Path(__file__).parent.parent / "metrics", help="Directory for snapshots and dashboard")
|
||||
args = parser.parse_args()
|
||||
|
||||
facts = load_facts(args.knowledge_dir)
|
||||
yaml_counts = count_yaml_facts(args.knowledge_dir)
|
||||
if args.repo:
|
||||
facts = [f for f in facts if f.get("domain") == args.repo]
|
||||
conn = open_db(args.db)
|
||||
sessions = query_sessions(conn, since=args.since)
|
||||
messages = query_messages(conn) if conn else []
|
||||
if conn:
|
||||
conn.close()
|
||||
|
||||
velocity = compute_knowledge_velocity(facts, since=args.since)
|
||||
coverage = compute_knowledge_coverage(facts, yaml_counts)
|
||||
hit_rate = compute_hit_rate(sessions, messages, facts)
|
||||
error_recurrence = compute_error_recurrence(messages)
|
||||
task_completion = compute_task_completion(sessions)
|
||||
first_try = compute_first_try_success(sessions)
|
||||
age = compute_knowledge_age(facts)
|
||||
|
||||
previous = load_snapshot(args.metrics_dir)
|
||||
trend = {
|
||||
"knowledge_velocity": compute_trend(velocity, previous.get("knowledge_velocity", {})),
|
||||
"knowledge_coverage": compute_trend(coverage, previous.get("knowledge_coverage", {})),
|
||||
"hit_rate": compute_trend(hit_rate, previous.get("hit_rate", {})),
|
||||
"error_recurrence": compute_trend(error_recurrence, previous.get("error_recurrence", {}), "value"),
|
||||
"task_completion": compute_trend(task_completion, previous.get("task_completion", {})),
|
||||
"first_try_success": compute_trend(first_try, previous.get("first_try_success", {})),
|
||||
"knowledge_age": compute_trend(age, previous.get("knowledge_age", {})),
|
||||
}
|
||||
|
||||
now = datetime.now(timezone.utc).isoformat()
|
||||
metrics = {
|
||||
"generated_at": now,
|
||||
"knowledge_velocity": velocity,
|
||||
"knowledge_coverage": coverage,
|
||||
"hit_rate": hit_rate,
|
||||
"error_recurrence": error_recurrence,
|
||||
"task_completion": task_completion,
|
||||
"first_try_success": first_try,
|
||||
"knowledge_age": age,
|
||||
"trend_7d": trend,
|
||||
}
|
||||
if args.since:
|
||||
metrics["since"] = args.since
|
||||
|
||||
if args.save_snapshot:
|
||||
save_snapshot(args.metrics_dir, metrics)
|
||||
dashboard_path = args.metrics_dir / "dashboard.md"
|
||||
with open(dashboard_path, "w") as f:
|
||||
f.write(format_markdown(metrics))
|
||||
|
||||
if args.format == "json":
|
||||
print(format_json(metrics))
|
||||
else:
|
||||
print(format_markdown(metrics))
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -1,239 +0,0 @@
|
||||
#!/usr/bin/env python3
|
||||
"""
|
||||
Tests for bootstrapper.py — context assembly from knowledge store.
|
||||
"""
|
||||
|
||||
import json
|
||||
import sys
|
||||
import tempfile
|
||||
from pathlib import Path
|
||||
|
||||
# Add scripts dir to path for import
|
||||
sys.path.insert(0, str(Path(__file__).resolve().parent))
|
||||
|
||||
from bootstrapper import (
|
||||
build_bootstrap_context,
|
||||
estimate_tokens,
|
||||
filter_facts,
|
||||
load_index,
|
||||
sort_facts,
|
||||
truncate_to_tokens,
|
||||
)
|
||||
|
||||
|
||||
def make_index(facts: list[dict], tmp_dir: Path) -> Path:
|
||||
"""Create a temporary index.json with given facts."""
|
||||
index = {
|
||||
"version": 1,
|
||||
"last_updated": "2026-04-13T20:00:00Z",
|
||||
"total_facts": len(facts),
|
||||
"facts": facts,
|
||||
}
|
||||
path = tmp_dir / "index.json"
|
||||
with open(path, "w") as f:
|
||||
json.dump(index, f)
|
||||
return path
|
||||
|
||||
|
||||
def test_empty_index():
|
||||
"""Empty knowledge store produces graceful output."""
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
tmp_dir = Path(tmp)
|
||||
index_path = make_index([], tmp_dir)
|
||||
|
||||
# Create empty knowledge dirs
|
||||
for sub in ["repos", "agents", "global"]:
|
||||
(tmp_dir / sub).mkdir(exist_ok=True)
|
||||
|
||||
context = build_bootstrap_context(
|
||||
repo="the-nexus", index_path=index_path
|
||||
)
|
||||
assert "No relevant knowledge found" in context
|
||||
assert "Starting fresh" in context
|
||||
print("PASS: empty_index")
|
||||
|
||||
|
||||
def test_filter_by_repo():
|
||||
"""Filter facts by repository."""
|
||||
facts = [
|
||||
{"fact": "A", "category": "fact", "repo": "the-nexus", "confidence": 0.9},
|
||||
{"fact": "B", "category": "fact", "repo": "fleet-ops", "confidence": 0.8},
|
||||
{"fact": "C", "category": "fact", "repo": "global", "confidence": 0.7},
|
||||
]
|
||||
filtered = filter_facts(facts, repo="the-nexus", include_global=True)
|
||||
texts = [f["fact"] for f in filtered]
|
||||
assert "A" in texts
|
||||
assert "B" not in texts
|
||||
assert "C" in texts
|
||||
print("PASS: filter_by_repo")
|
||||
|
||||
|
||||
def test_filter_by_agent():
|
||||
"""Filter facts by agent type."""
|
||||
facts = [
|
||||
{"fact": "A", "category": "pattern", "repo": "global", "agent": "mimo-sprint", "confidence": 0.8},
|
||||
{"fact": "B", "category": "pattern", "repo": "global", "agent": "groq-fast", "confidence": 0.7},
|
||||
{"fact": "C", "category": "fact", "repo": "global", "confidence": 0.9},
|
||||
]
|
||||
filtered = filter_facts(facts, agent="mimo-sprint", include_global=True)
|
||||
texts = [f["fact"] for f in filtered]
|
||||
assert "A" in texts
|
||||
assert "B" not in texts
|
||||
assert "C" in texts # global, no agent restriction
|
||||
print("PASS: filter_by_agent")
|
||||
|
||||
|
||||
def test_no_global_flag():
|
||||
"""Excluding global facts works."""
|
||||
facts = [
|
||||
{"fact": "A", "category": "fact", "repo": "the-nexus", "confidence": 0.9},
|
||||
{"fact": "B", "category": "fact", "repo": "global", "confidence": 0.8},
|
||||
]
|
||||
filtered = filter_facts(facts, repo="the-nexus", include_global=False)
|
||||
texts = [f["fact"] for f in filtered]
|
||||
assert "A" in texts
|
||||
assert "B" not in texts
|
||||
print("PASS: no_global_flag")
|
||||
|
||||
|
||||
def test_sort_by_confidence():
|
||||
"""Facts sort by confidence descending."""
|
||||
facts = [
|
||||
{"fact": "low", "category": "fact", "repo": "global", "confidence": 0.3},
|
||||
{"fact": "high", "category": "fact", "repo": "global", "confidence": 0.95},
|
||||
{"fact": "mid", "category": "fact", "repo": "global", "confidence": 0.7},
|
||||
]
|
||||
sorted_f = sort_facts(facts)
|
||||
assert sorted_f[0]["fact"] == "high"
|
||||
assert sorted_f[1]["fact"] == "mid"
|
||||
assert sorted_f[2]["fact"] == "low"
|
||||
print("PASS: sort_by_confidence")
|
||||
|
||||
|
||||
def test_sort_pitfalls_first():
|
||||
"""Pitfalls sort before facts at same confidence."""
|
||||
facts = [
|
||||
{"fact": "regular fact", "category": "fact", "repo": "global", "confidence": 0.8},
|
||||
{"fact": "danger pitfall", "category": "pitfall", "repo": "global", "confidence": 0.8},
|
||||
]
|
||||
sorted_f = sort_facts(facts)
|
||||
assert sorted_f[0]["category"] == "pitfall"
|
||||
print("PASS: sort_pitfalls_first")
|
||||
|
||||
|
||||
def test_truncate_to_tokens():
|
||||
"""Truncation cuts at line boundary."""
|
||||
text = "line1\nline2\nline3\nline4\nline5\n"
|
||||
truncated = truncate_to_tokens(text, max_tokens=2) # ~8 chars
|
||||
assert "line1" in truncated
|
||||
assert "truncated" in truncated.lower()
|
||||
print("PASS: truncate_to_tokens")
|
||||
|
||||
|
||||
def test_estimate_tokens():
|
||||
"""Token estimation is reasonable."""
|
||||
text = "a" * 400
|
||||
tokens = estimate_tokens(text)
|
||||
assert 90 <= tokens <= 110 # ~100 tokens
|
||||
print("PASS: estimate_tokens")
|
||||
|
||||
|
||||
def test_build_full_context():
|
||||
"""Full context with facts renders correctly."""
|
||||
facts = [
|
||||
{"fact": "API merges fail with 405", "category": "pitfall", "repo": "the-nexus", "confidence": 0.95},
|
||||
{"fact": "Has 50+ open PRs", "category": "fact", "repo": "the-nexus", "confidence": 0.9},
|
||||
{"fact": "Token at ~/.config/gitea/token", "category": "tool-quirk", "repo": "global", "confidence": 0.9},
|
||||
{"fact": "Check git remote -v first", "category": "pattern", "repo": "global", "confidence": 0.8},
|
||||
]
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
tmp_dir = Path(tmp)
|
||||
index_path = make_index(facts, tmp_dir)
|
||||
|
||||
# Create knowledge dirs
|
||||
for sub in ["repos", "agents", "global"]:
|
||||
(tmp_dir / sub).mkdir(exist_ok=True)
|
||||
|
||||
context = build_bootstrap_context(
|
||||
repo="the-nexus",
|
||||
agent="mimo-sprint",
|
||||
include_global=True,
|
||||
index_path=index_path,
|
||||
)
|
||||
|
||||
assert "What You Know" in context
|
||||
assert "PITFALLS" in context
|
||||
assert "API merges fail with 405" in context
|
||||
assert "the-nexus" in context
|
||||
assert "Token at" in context # global fact included
|
||||
print("PASS: build_full_context")
|
||||
|
||||
|
||||
def test_max_tokens_respected():
|
||||
"""Output respects max_tokens limit."""
|
||||
# Generate lots of facts
|
||||
facts = [
|
||||
{"fact": f"Fact number {i} with some detail about things", "category": "fact", "repo": "global", "confidence": 0.8}
|
||||
for i in range(100)
|
||||
]
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
tmp_dir = Path(tmp)
|
||||
index_path = make_index(facts, tmp_dir)
|
||||
for sub in ["repos", "agents", "global"]:
|
||||
(tmp_dir / sub).mkdir(exist_ok=True)
|
||||
|
||||
context = build_bootstrap_context(
|
||||
repo=None,
|
||||
max_tokens=500,
|
||||
index_path=index_path,
|
||||
)
|
||||
|
||||
actual_tokens = estimate_tokens(context)
|
||||
# Allow 10% overshoot since we cut at line boundaries
|
||||
assert actual_tokens <= 550, f"Expected ~500 tokens, got {actual_tokens}"
|
||||
print(f"PASS: max_tokens_respected (got {actual_tokens} tokens)")
|
||||
|
||||
|
||||
def test_missing_index_graceful():
|
||||
"""Missing index.json doesn't crash."""
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
tmp_dir = Path(tmp)
|
||||
# Don't create index.json
|
||||
for sub in ["repos", "agents", "global"]:
|
||||
(tmp_dir / sub).mkdir(exist_ok=True)
|
||||
|
||||
fake_index = tmp_dir / "nonexistent.json"
|
||||
context = build_bootstrap_context(repo="anything", index_path=fake_index)
|
||||
assert "No relevant knowledge found" in context
|
||||
print("PASS: missing_index_graceful")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
tests = [
|
||||
test_empty_index,
|
||||
test_filter_by_repo,
|
||||
test_filter_by_agent,
|
||||
test_no_global_flag,
|
||||
test_sort_by_confidence,
|
||||
test_sort_pitfalls_first,
|
||||
test_truncate_to_tokens,
|
||||
test_estimate_tokens,
|
||||
test_build_full_context,
|
||||
test_max_tokens_respected,
|
||||
test_missing_index_graceful,
|
||||
]
|
||||
|
||||
passed = 0
|
||||
failed = 0
|
||||
for test in tests:
|
||||
try:
|
||||
test()
|
||||
passed += 1
|
||||
except Exception as e:
|
||||
print(f"FAIL: {test.__name__} — {e}")
|
||||
failed += 1
|
||||
|
||||
print(f"\n{passed} passed, {failed} failed")
|
||||
sys.exit(0 if failed == 0 else 1)
|
||||
@@ -1,129 +1,41 @@
|
||||
#!/usr/bin/env python3
|
||||
"""
|
||||
Test harness for knowledge extraction prompt.
|
||||
Validates output structure, content quality, and hallucination resistance.
|
||||
|
||||
Usage:
|
||||
python3 scripts/test_harvest_prompt.py # Run all tests
|
||||
python3 scripts/test_harvest_prompt.py --transcript FILE # Test against a real transcript
|
||||
python3 scripts/test_harvest_prompt.py --validate FILE # Validate an existing extraction JSON
|
||||
Test script for knowledge extraction prompt.
|
||||
Validates that the prompt produces consistent, structured output.
|
||||
"""
|
||||
|
||||
import json
|
||||
import sys
|
||||
import argparse
|
||||
from pathlib import Path
|
||||
|
||||
VALID_CATEGORIES = {"fact", "pitfall", "pattern", "tool-quirk", "question"}
|
||||
REQUIRED_FIELDS = {"fact", "category", "repo", "confidence", "evidence"}
|
||||
REQUIRED_META = {"session_outcome", "tools_used", "repos_touched", "error_count", "knowledge_count"}
|
||||
|
||||
|
||||
def validate_knowledge_item(item, idx):
|
||||
"""Validate a single knowledge item. Returns list of errors."""
|
||||
errors = []
|
||||
if not isinstance(item, dict):
|
||||
return [f"Item {idx}: not a dict"]
|
||||
for field in REQUIRED_FIELDS:
|
||||
def validate_knowledge_item(item):
|
||||
"""Validate a single knowledge item."""
|
||||
required_fields = ["fact", "category", "repo", "confidence"]
|
||||
for field in required_fields:
|
||||
if field not in item:
|
||||
errors.append(f"Item {idx}: missing field '{field}'")
|
||||
if not isinstance(item.get("fact", ""), str) or len(item.get("fact", "").strip()) == 0:
|
||||
errors.append(f"Item {idx}: fact must be a non-empty string")
|
||||
if item.get("category") not in VALID_CATEGORIES:
|
||||
errors.append(f"Item {idx}: invalid category '{item.get('category')}'")
|
||||
if not isinstance(item.get("repo", ""), str) or len(item.get("repo", "").strip()) == 0:
|
||||
errors.append(f"Item {idx}: repo must be a non-empty string")
|
||||
conf = item.get("confidence")
|
||||
if not isinstance(conf, (int, float)) or not (0.0 <= conf <= 1.0):
|
||||
errors.append(f"Item {idx}: confidence must be a number 0.0-1.0, got {conf}")
|
||||
if not isinstance(item.get("evidence", ""), str) or len(item.get("evidence", "").strip()) == 0:
|
||||
errors.append(f"Item {idx}: evidence must be a non-empty string (hallucination check)")
|
||||
return errors
|
||||
return False, f"Missing field: {field}"
|
||||
|
||||
if not isinstance(item["fact"], str) or len(item["fact"].strip()) == 0:
|
||||
return False, "Fact must be a non-empty string"
|
||||
|
||||
valid_categories = ["fact", "pitfall", "pattern", "tool-quirk", "question"]
|
||||
if item["category"] not in valid_categories:
|
||||
return False, f"Invalid category: {item['category']}"
|
||||
|
||||
if not isinstance(item["repo"], str):
|
||||
return False, "Repo must be a string"
|
||||
|
||||
if not isinstance(item["confidence"], (int, float)):
|
||||
return False, "Confidence must be a number"
|
||||
|
||||
if not (0.0 <= item["confidence"] <= 1.0):
|
||||
return False, "Confidence must be between 0.0 and 1.0"
|
||||
|
||||
return True, "Valid"
|
||||
|
||||
|
||||
def validate_extraction(data):
|
||||
"""Validate a full extraction result. Returns (is_valid, errors, warnings)."""
|
||||
errors = []
|
||||
warnings = []
|
||||
|
||||
if not isinstance(data, dict):
|
||||
return False, ["Root is not a JSON object"], []
|
||||
|
||||
if "knowledge" not in data:
|
||||
return False, ["Missing 'knowledge' array"], []
|
||||
|
||||
if not isinstance(data["knowledge"], list):
|
||||
return False, ["'knowledge' is not an array"], []
|
||||
|
||||
for i, item in enumerate(data["knowledge"]):
|
||||
errors.extend(validate_knowledge_item(item, i))
|
||||
|
||||
# Meta block validation
|
||||
if "meta" not in data:
|
||||
warnings.append("Missing 'meta' block (session_outcome, tools_used, etc.)")
|
||||
else:
|
||||
meta = data["meta"]
|
||||
for field in REQUIRED_META:
|
||||
if field not in meta:
|
||||
warnings.append(f"Meta missing field '{field}'")
|
||||
|
||||
# Quality checks
|
||||
facts = data["knowledge"]
|
||||
if len(facts) == 0:
|
||||
warnings.append("No knowledge extracted (empty session or extraction failure)")
|
||||
|
||||
# Check for near-duplicate facts
|
||||
seen_facts = set()
|
||||
for item in facts:
|
||||
normalized = item.get("fact", "").lower().strip()[:80]
|
||||
if normalized in seen_facts:
|
||||
warnings.append(f"Duplicate fact detected: '{normalized[:50]}...'")
|
||||
seen_facts.add(normalized)
|
||||
|
||||
# Check confidence distribution
|
||||
confidences = [item.get("confidence", 0) for item in facts]
|
||||
if confidences:
|
||||
avg_conf = sum(confidences) / len(confidences)
|
||||
if avg_conf > 0.9:
|
||||
warnings.append(f"Average confidence {avg_conf:.2f} is suspiciously high (possible hallucination)")
|
||||
if avg_conf < 0.4:
|
||||
warnings.append(f"Average confidence {avg_conf:.2f} is very low (extraction may be too cautious)")
|
||||
|
||||
return len(errors) == 0, errors, warnings
|
||||
|
||||
|
||||
def validate_transcript_coverage(data, transcript):
|
||||
"""
|
||||
Check that extracted facts are actually supported by the transcript.
|
||||
This is a heuristic — checks that key terms from facts appear in transcript.
|
||||
Returns list of potential hallucinations.
|
||||
"""
|
||||
hallucinations = []
|
||||
transcript_lower = transcript.lower()
|
||||
|
||||
for item in data.get("knowledge", []):
|
||||
fact = item.get("fact", "")
|
||||
evidence = item.get("evidence", "")
|
||||
|
||||
# Check if evidence string appears in transcript
|
||||
if evidence and evidence.lower() not in transcript_lower:
|
||||
# Partial match — check if key terms are present
|
||||
evidence_words = set(evidence.lower().split())
|
||||
transcript_words = set(transcript_lower.split())
|
||||
overlap = evidence_words & transcript_words
|
||||
if len(overlap) < min(3, len(evidence_words) * 0.3):
|
||||
hallucinations.append({
|
||||
"fact": fact[:80],
|
||||
"evidence": evidence[:80],
|
||||
"issue": "Evidence not found in transcript — possible hallucination"
|
||||
})
|
||||
|
||||
return hallucinations
|
||||
|
||||
|
||||
# --- Test transcripts ---
|
||||
|
||||
SAMPLE_TRANSCRIPT_SUCCESS = """
|
||||
def test_sample_transcript():
|
||||
"""Test with a sample transcript."""
|
||||
sample_transcript = """
|
||||
User: Clone the repo at forge.alexanderwhitestone.com/Timmy_Foundation/compounding-intelligence
|
||||
Agent: Trying git clone... Error: 401 Unauthorized
|
||||
Agent: Checking for token... Found in ~/.config/gitea/token
|
||||
@@ -132,253 +44,30 @@ Agent: Clone successful with token authentication
|
||||
Agent: Now examining the repo structure...
|
||||
Agent: Found knowledge/index.json with 0 facts
|
||||
Agent: Creating harvest-prompt.md in templates/
|
||||
Agent: Done. Files created: templates/harvest-prompt.md, scripts/test_harvest_prompt.py
|
||||
"""
|
||||
|
||||
SAMPLE_TRANSCRIPT_FAILURE = """
|
||||
User: Deploy the cron jobs to production
|
||||
Agent: Connecting to VPS... SSH timeout after 30s
|
||||
Agent: Retrying with verbose output... Connection refused on port 22
|
||||
Agent: Checking firewall rules... iptables shows port 22 is open
|
||||
Agent: Checking sshd config... sshd is not running
|
||||
Agent: Attempting to start sshd... systemctl start sshd failed: unit not found
|
||||
Agent: ERROR: OpenSSH server not installed on this VPS
|
||||
Agent: Cannot proceed without SSH access. Deployment blocked.
|
||||
"""
|
||||
|
||||
SAMPLE_TRANSCRIPT_API_QUIRK = """
|
||||
User: List all open issues in hermes-agent
|
||||
Agent: Querying Gitea API... endpoint: /repos/Timmy_Foundation/hermes-agent/issues?state=open&limit=50
|
||||
Agent: Response: 0 items. That's odd.
|
||||
Agent: Trying without state filter... endpoint: /repos/Timmy_Foundation/hermes-agent/issues?limit=50
|
||||
Agent: Response: 50 items returned, but all have pull_request field. These are PRs, not issues.
|
||||
Agent: Gitea issues API returns PRs alongside issues. Must filter by presence of pull_request key.
|
||||
Agent: Also discovered: state=open parameter returns empty when used with limit=50, but works with limit=5.
|
||||
Agent: This appears to be a Gitea API quirk — pagination interacts with state filter unexpectedly.
|
||||
Agent: Workaround: fetch without state filter, then filter client-side for state=='open' and 'pull_request' not in item.
|
||||
"""
|
||||
|
||||
|
||||
def run_tests():
|
||||
"""Run the built-in test suite."""
|
||||
tests_passed = 0
|
||||
tests_failed = 0
|
||||
|
||||
print("=" * 60)
|
||||
print("KNOWLEDGE EXTRACTION PROMPT — TEST SUITE")
|
||||
print("=" * 60)
|
||||
|
||||
# Test 1: Prompt file exists and is under 2k tokens (~8k chars)
|
||||
print("\n[Test 1] Prompt file size constraint")
|
||||
prompt_path = Path("templates/harvest-prompt.md")
|
||||
if not prompt_path.exists():
|
||||
print(" FAIL: harvest-prompt.md not found")
|
||||
tests_failed += 1
|
||||
else:
|
||||
size = prompt_path.stat().st_size
|
||||
# Rough token estimate: ~4 chars per token
|
||||
est_tokens = size / 4
|
||||
print(f" Prompt size: {size} bytes (~{est_tokens:.0f} tokens)")
|
||||
if est_tokens > 2000:
|
||||
print(f" WARN: Prompt exceeds ~1500 tokens (target: ~1000)")
|
||||
else:
|
||||
print(f" PASS: Within token budget")
|
||||
tests_passed += 1
|
||||
|
||||
# Test 2: Validate a well-formed extraction
|
||||
print("\n[Test 2] Valid extraction passes validation")
|
||||
valid_extraction = {
|
||||
"knowledge": [
|
||||
{
|
||||
"fact": "Gitea auth token is at ~/.config/gitea/token",
|
||||
"category": "tool-quirk",
|
||||
"repo": "global",
|
||||
"confidence": 0.9,
|
||||
"evidence": "Found in ~/.config/gitea/token"
|
||||
},
|
||||
{
|
||||
"fact": "Clone fails with 401 when no token is provided",
|
||||
"category": "pitfall",
|
||||
"repo": "compounding-intelligence",
|
||||
"confidence": 0.9,
|
||||
"evidence": "Error: 401 Unauthorized"
|
||||
}
|
||||
],
|
||||
"meta": {
|
||||
"session_outcome": "success",
|
||||
"tools_used": ["git"],
|
||||
"repos_touched": ["compounding-intelligence"],
|
||||
"error_count": 1,
|
||||
"knowledge_count": 2
|
||||
}
|
||||
}
|
||||
is_valid, errors, warnings = validate_extraction(valid_extraction)
|
||||
if is_valid:
|
||||
print(f" PASS: Valid extraction accepted ({len(warnings)} warnings)")
|
||||
tests_passed += 1
|
||||
else:
|
||||
print(f" FAIL: Valid extraction rejected: {errors}")
|
||||
tests_failed += 1
|
||||
|
||||
# Test 3: Reject missing fields
|
||||
print("\n[Test 3] Missing fields are rejected")
|
||||
bad_extraction = {
|
||||
"knowledge": [
|
||||
{"fact": "Something learned", "category": "fact"} # Missing repo, confidence, evidence
|
||||
]
|
||||
}
|
||||
is_valid, errors, warnings = validate_extraction(bad_extraction)
|
||||
if not is_valid:
|
||||
print(f" PASS: Rejected with {len(errors)} errors")
|
||||
tests_passed += 1
|
||||
else:
|
||||
print(f" FAIL: Should have rejected missing fields")
|
||||
tests_failed += 1
|
||||
|
||||
# Test 4: Reject invalid category
|
||||
print("\n[Test 4] Invalid category is rejected")
|
||||
bad_cat = {
|
||||
"knowledge": [
|
||||
{"fact": "Test", "category": "discovery", "repo": "x", "confidence": 0.8, "evidence": "test"}
|
||||
]
|
||||
}
|
||||
is_valid, errors, warnings = validate_extraction(bad_cat)
|
||||
if not is_valid and any("category" in e for e in errors):
|
||||
print(f" PASS: Invalid category 'discovery' rejected")
|
||||
tests_passed += 1
|
||||
else:
|
||||
print(f" FAIL: Should have rejected invalid category")
|
||||
tests_failed += 1
|
||||
|
||||
# Test 5: Detect near-duplicates
|
||||
print("\n[Test 5] Duplicate detection")
|
||||
dup_extraction = {
|
||||
"knowledge": [
|
||||
{"fact": "Token is at ~/.config/gitea/token", "category": "fact", "repo": "x", "confidence": 0.9, "evidence": "a"},
|
||||
{"fact": "Token is at ~/.config/gitea/token", "category": "fact", "repo": "x", "confidence": 0.9, "evidence": "b"}
|
||||
],
|
||||
"meta": {"session_outcome": "success", "tools_used": [], "repos_touched": [], "error_count": 0, "knowledge_count": 2}
|
||||
}
|
||||
is_valid, errors, warnings = validate_extraction(dup_extraction)
|
||||
if any("Duplicate" in w for w in warnings):
|
||||
print(f" PASS: Duplicate detected")
|
||||
tests_passed += 1
|
||||
else:
|
||||
print(f" FAIL: Should have detected duplicate")
|
||||
tests_failed += 1
|
||||
|
||||
# Test 6: Hallucination check against transcript
|
||||
print("\n[Test 6] Hallucination detection")
|
||||
hallucinated = {
|
||||
"knowledge": [
|
||||
{
|
||||
"fact": "Database port is 5433",
|
||||
"category": "fact",
|
||||
"repo": "x",
|
||||
"confidence": 0.9,
|
||||
"evidence": "PostgreSQL listening on port 5433"
|
||||
}
|
||||
],
|
||||
"meta": {"session_outcome": "success", "tools_used": [], "repos_touched": [], "error_count": 0, "knowledge_count": 1}
|
||||
}
|
||||
hallucinations = validate_transcript_coverage(hallucinated, SAMPLE_TRANSCRIPT_SUCCESS)
|
||||
if hallucinations:
|
||||
print(f" PASS: Hallucination detected ({len(hallucinations)} items)")
|
||||
tests_passed += 1
|
||||
else:
|
||||
print(f" FAIL: Should have detected hallucinated evidence")
|
||||
tests_failed += 1
|
||||
|
||||
# Test 7: Failed session should extract pitfalls
|
||||
print("\n[Test 7] Failed session extraction shape")
|
||||
failed_extraction = {
|
||||
"knowledge": [
|
||||
{
|
||||
"fact": "SSH server not installed on target VPS",
|
||||
"category": "pitfall",
|
||||
"repo": "global",
|
||||
"confidence": 0.9,
|
||||
"evidence": "ERROR: OpenSSH server not installed on this VPS"
|
||||
},
|
||||
{
|
||||
"fact": "VPS blocks deployment without SSH access",
|
||||
"category": "question",
|
||||
"repo": "global",
|
||||
"confidence": 0.7,
|
||||
"evidence": "Cannot proceed without SSH access. Deployment blocked."
|
||||
}
|
||||
],
|
||||
"meta": {
|
||||
"session_outcome": "failed",
|
||||
"tools_used": ["ssh", "systemctl"],
|
||||
"repos_touched": [],
|
||||
"error_count": 3,
|
||||
"knowledge_count": 2
|
||||
}
|
||||
}
|
||||
is_valid, errors, warnings = validate_extraction(failed_extraction)
|
||||
if is_valid:
|
||||
categories = [item["category"] for item in failed_extraction["knowledge"]]
|
||||
if "pitfall" in categories:
|
||||
print(f" PASS: Failed session extracted {len(categories)} items including pitfalls")
|
||||
tests_passed += 1
|
||||
else:
|
||||
print(f" FAIL: Failed session should extract pitfalls")
|
||||
tests_failed += 1
|
||||
else:
|
||||
print(f" FAIL: {errors}")
|
||||
tests_failed += 1
|
||||
|
||||
# Test 8: Empty extraction is warned
|
||||
print("\n[Test 8] Empty extraction warning")
|
||||
empty = {"knowledge": [], "meta": {"session_outcome": "success", "tools_used": [], "repos_touched": [], "error_count": 0, "knowledge_count": 0}}
|
||||
is_valid, errors, warnings = validate_extraction(empty)
|
||||
if any("No knowledge" in w for w in warnings):
|
||||
print(f" PASS: Empty extraction warned")
|
||||
tests_passed += 1
|
||||
else:
|
||||
print(f" FAIL: Should warn on empty extraction")
|
||||
tests_failed += 1
|
||||
|
||||
# Summary
|
||||
print(f"\n{'=' * 60}")
|
||||
print(f"Results: {tests_passed} passed, {tests_failed} failed")
|
||||
print(f"{'=' * 60}")
|
||||
return tests_failed == 0
|
||||
|
||||
|
||||
def validate_file(filepath):
|
||||
"""Validate an existing extraction JSON file."""
|
||||
path = Path(filepath)
|
||||
if not path.exists():
|
||||
print(f"ERROR: {filepath} not found")
|
||||
return False
|
||||
|
||||
data = json.loads(path.read_text())
|
||||
is_valid, errors, warnings = validate_extraction(data)
|
||||
|
||||
print(f"Validation of {filepath}:")
|
||||
print(f" Knowledge items: {len(data.get('knowledge', []))}")
|
||||
print(f" Errors: {len(errors)}")
|
||||
print(f" Warnings: {len(warnings)}")
|
||||
|
||||
for e in errors:
|
||||
print(f" ERROR: {e}")
|
||||
for w in warnings:
|
||||
print(f" WARN: {w}")
|
||||
|
||||
return is_valid
|
||||
|
||||
|
||||
# This would be replaced with actual prompt execution
|
||||
print("Sample transcript processed")
|
||||
print("Expected categories: fact, pitfall, pattern, tool-quirk, question")
|
||||
return True
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser(description="Test knowledge extraction prompt")
|
||||
parser.add_argument("--validate", help="Validate an existing extraction JSON file")
|
||||
parser.add_argument("--transcript", help="Test against a real transcript file (informational)")
|
||||
args = parser.parse_args()
|
||||
|
||||
if args.validate:
|
||||
success = validate_file(args.validate)
|
||||
sys.exit(0 if success else 1)
|
||||
else:
|
||||
success = run_tests()
|
||||
sys.exit(0 if success else 1)
|
||||
print("Testing knowledge extraction prompt...")
|
||||
|
||||
# Test 1: Validate prompt file exists
|
||||
prompt_path = Path("templates/harvest-prompt.md")
|
||||
if not prompt_path.exists():
|
||||
print("ERROR: harvest-prompt.md not found")
|
||||
sys.exit(1)
|
||||
|
||||
print(f"OK: Prompt file exists: {prompt_path}")
|
||||
|
||||
# Test 2: Check prompt size
|
||||
prompt_size = prompt_path.stat().st_size
|
||||
print(f"OK: Prompt size: {prompt_size} bytes")
|
||||
|
||||
# Test 3: Test sample transcript processing
|
||||
if test_sample_transcript():
|
||||
print("OK: Sample transcript test passed")
|
||||
|
||||
print("\nAll tests passed!")
|
||||
|
||||
80
scripts/validate_knowledge.py
Normal file
80
scripts/validate_knowledge.py
Normal file
@@ -0,0 +1,80 @@
|
||||
#!/usr/bin/env python3
|
||||
"""
|
||||
Validate knowledge files and index.json against the schema.
|
||||
|
||||
Usage:
|
||||
python scripts/validate_knowledge.py
|
||||
"""
|
||||
|
||||
import json
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
VALID_CATEGORIES = {"fact", "pitfall", "pattern", "tool-quirk", "question"}
|
||||
REQUIRED_FACT_FIELDS = {"id", "fact", "category", "domain", "confidence"}
|
||||
MAX_FACT_LENGTH = 280
|
||||
|
||||
|
||||
def validate_fact(fact, source=""):
|
||||
errors = []
|
||||
for field in REQUIRED_FACT_FIELDS:
|
||||
if field not in fact:
|
||||
errors.append(f"{source}: missing required field '{field}'")
|
||||
if "fact" in fact:
|
||||
if not isinstance(fact["fact"], str) or len(fact["fact"].strip()) == 0:
|
||||
errors.append(f"{source}: 'fact' must be non-empty string")
|
||||
elif len(fact["fact"]) > MAX_FACT_LENGTH:
|
||||
errors.append(f"{source}: 'fact' exceeds {MAX_FACT_LENGTH} chars")
|
||||
if "category" in fact and fact["category"] not in VALID_CATEGORIES:
|
||||
errors.append(f"{source}: invalid category '{fact['category']}'")
|
||||
if "confidence" in fact:
|
||||
if not isinstance(fact["confidence"], (int, float)):
|
||||
errors.append(f"{source}: 'confidence' must be a number")
|
||||
elif not (0.0 <= fact["confidence"] <= 1.0):
|
||||
errors.append(f"{source}: 'confidence' must be 0.0-1.0")
|
||||
if "id" in fact:
|
||||
parts = fact["id"].split(":")
|
||||
if len(parts) != 3:
|
||||
errors.append(f"{source}: 'id' must be domain:category:sequence")
|
||||
elif parts[1] not in VALID_CATEGORIES:
|
||||
errors.append(f"{source}: id category '{parts[1]}' invalid")
|
||||
return errors
|
||||
|
||||
|
||||
def main():
|
||||
repo_root = Path(__file__).parent.parent
|
||||
index_path = repo_root / "knowledge" / "index.json"
|
||||
all_errors = []
|
||||
|
||||
if not index_path.exists():
|
||||
print(f"VALIDATION FAILED: index.json not found at {index_path}")
|
||||
sys.exit(1)
|
||||
|
||||
with open(index_path) as f:
|
||||
data = json.load(f)
|
||||
|
||||
if "version" not in data:
|
||||
all_errors.append("index.json: missing 'version'")
|
||||
if "facts" not in data or not isinstance(data["facts"], list):
|
||||
all_errors.append("index.json: missing or invalid 'facts'")
|
||||
|
||||
seen_ids = set()
|
||||
for i, fact in enumerate(data.get("facts", [])):
|
||||
all_errors.extend(validate_fact(fact, f"facts[{i}]"))
|
||||
if "id" in fact:
|
||||
if fact["id"] in seen_ids:
|
||||
all_errors.append(f"index.json: duplicate id '{fact['id']}'")
|
||||
seen_ids.add(fact["id"])
|
||||
|
||||
if all_errors:
|
||||
print(f"VALIDATION FAILED - {len(all_errors)} error(s):\n")
|
||||
for e in all_errors:
|
||||
print(f" x {e}")
|
||||
sys.exit(1)
|
||||
else:
|
||||
print(f"VALIDATION PASSED - {len(data.get('facts', []))} facts, schema v{data.get('version', '?')}")
|
||||
sys.exit(0)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -2,107 +2,98 @@
|
||||
|
||||
## System Prompt
|
||||
|
||||
You are a knowledge extraction engine. You read session transcripts and output ONLY structured JSON. You never infer. You never assume. You extract only what the transcript explicitly states.
|
||||
You are a knowledge extraction engine. Your task is to analyze a session transcript and extract durable knowledge that will help future sessions be more efficient.
|
||||
|
||||
## Prompt
|
||||
## Instructions
|
||||
|
||||
Read the session transcript carefully. Extract ONLY information that is explicitly stated in the transcript. Do NOT infer, assume, or hallucinate information.
|
||||
|
||||
### Categories
|
||||
|
||||
Extract knowledge into these categories:
|
||||
|
||||
1. **fact**: Concrete, verifiable information learned (e.g., "Repository X has 5 files", "API returns JSON with field Y")
|
||||
2. **pitfall**: Errors encountered, wrong assumptions, things that wasted time (e.g., "Assumed API token was in env var GITEA_TOKEN, but it's in ~/.config/gitea/token")
|
||||
3. **pattern**: Successful sequences of actions (e.g., "To deploy: 1. Run tests 2. Build 3. Push to Gitea 4. Trigger webhook")
|
||||
4. **tool-quirk**: Environment-specific behaviors (e.g., "Token paths are different on macOS vs Linux", "URL format requires trailing slash")
|
||||
5. **question**: Things identified but not answered (e.g., "Need to determine optimal batch size for harvesting")
|
||||
|
||||
### Output Format
|
||||
|
||||
Return a JSON object with an array of extracted knowledge items. Each item must have:
|
||||
|
||||
```json
|
||||
{
|
||||
"fact": "One sentence description of the knowledge",
|
||||
"category": "fact|pitfall|pattern|tool-quirk|question",
|
||||
"repo": "Repository name this applies to, or 'global' if general",
|
||||
"confidence": 0.0-1.0
|
||||
}
|
||||
```
|
||||
TASK: Extract durable knowledge from this session transcript.
|
||||
|
||||
RULES:
|
||||
1. Extract ONLY information explicitly stated in the transcript.
|
||||
2. Do NOT infer, assume, or hallucinate.
|
||||
3. Every fact must be verifiable by pointing to a specific line in the transcript.
|
||||
4. If the session failed or was partial, extract pitfalls and questions — these are the most valuable.
|
||||
5. Be specific. "Gitea API is slow" is worthless. "Gitea issues endpoint with state=open returns empty when limit=50 but works with limit=5" is knowledge.
|
||||
### Confidence Scoring
|
||||
|
||||
CATEGORIES (assign exactly one per item):
|
||||
- fact: Concrete, verifiable thing learned (paths, formats, counts, configs)
|
||||
- pitfall: Error hit, wrong assumption, time wasted, thing that didn't work
|
||||
- pattern: Successful sequence that should be reused (deploy steps, debug flow)
|
||||
- tool-quirk: Environment-specific behavior (token paths, URL formats, API gotchas)
|
||||
- question: Something identified but not answered — the NEXT agent should investigate
|
||||
- 0.9-1.0: Explicitly stated with verification (e.g., "Error message shows X")
|
||||
- 0.7-0.8: Clearly implied by multiple data points
|
||||
- 0.5-0.6: Suggested but not fully verified
|
||||
- 0.3-0.4: Inferred from limited data
|
||||
- 0.1-0.2: Speculative or uncertain
|
||||
|
||||
CONFIDENCE:
|
||||
- 0.9: Directly observed with error output or explicit verification
|
||||
- 0.7: Multiple data points confirm, but not explicitly verified
|
||||
- 0.5: Suggested by context, not tested
|
||||
- 0.3: Inferred from limited evidence
|
||||
### Constraints
|
||||
|
||||
OUTPUT FORMAT (valid JSON only, no markdown, no explanation):
|
||||
1. **No hallucination**: Only extract what's explicitly in the transcript
|
||||
2. **Specificity**: Each fact must be specific and actionable
|
||||
3. **Relevance**: Only extract knowledge that would help future sessions
|
||||
4. **Brevity**: One sentence per fact
|
||||
5. **Partial sessions**: Even failed or incomplete sessions may contain valuable pitfalls
|
||||
|
||||
### Example Input/Output
|
||||
|
||||
**Input Transcript (excerpt):**
|
||||
```
|
||||
User: Clone the repo at forge.alexanderwhitestone.com/Timmy_Foundation/compounding-intelligence
|
||||
Agent: Trying git clone... Error: 401 Unauthorized
|
||||
Agent: Checking for token... Found in ~/.config/gitea/token
|
||||
Agent: Token is gitea_token format, not OAuth
|
||||
Agent: Clone successful with token authentication
|
||||
```
|
||||
|
||||
**Output:**
|
||||
```json
|
||||
{
|
||||
"knowledge": [
|
||||
{
|
||||
"fact": "One specific sentence of knowledge",
|
||||
"category": "fact|pitfall|pattern|tool-quirk|question",
|
||||
"repo": "repo-name or global",
|
||||
"confidence": 0.0-1.0,
|
||||
"evidence": "Brief quote or reference from transcript that supports this"
|
||||
"fact": "Gitea repo at forge.alexanderwhitestone.com requires authentication for cloning",
|
||||
"category": "fact",
|
||||
"repo": "compounding-intelligence",
|
||||
"confidence": 0.9
|
||||
},
|
||||
{
|
||||
"fact": "Gitea authentication token is stored at ~/.config/gitea/token",
|
||||
"category": "tool-quirk",
|
||||
"repo": "global",
|
||||
"confidence": 0.9
|
||||
},
|
||||
{
|
||||
"fact": "Gitea uses gitea_token format, not OAuth for API access",
|
||||
"category": "tool-quirk",
|
||||
"repo": "global",
|
||||
"confidence": 0.8
|
||||
},
|
||||
{
|
||||
"fact": "Clone fails with 401 when no token is provided",
|
||||
"category": "pitfall",
|
||||
"repo": "compounding-intelligence",
|
||||
"confidence": 0.9
|
||||
}
|
||||
],
|
||||
"meta": {
|
||||
"session_outcome": "success|partial|failed",
|
||||
"tools_used": ["tool1", "tool2"],
|
||||
"repos_touched": ["repo1"],
|
||||
"error_count": 0,
|
||||
"knowledge_count": 0
|
||||
}
|
||||
]
|
||||
}
|
||||
|
||||
TRANSCRIPT:
|
||||
{{transcript}}
|
||||
```
|
||||
|
||||
## Design Notes
|
||||
## Final Notes
|
||||
|
||||
### Why this works with mimo-v2-pro
|
||||
|
||||
Mimo needs:
|
||||
- Explicit format constraints ("valid JSON only, no markdown")
|
||||
- Clear category definitions with concrete examples
|
||||
- Hard rules before soft guidance
|
||||
- The transcript at the END (so it reads all instructions first)
|
||||
|
||||
This prompt front-loads all rules, then gives the transcript last. Mimo follows the pattern.
|
||||
|
||||
### Handling partial/failed sessions
|
||||
|
||||
Failed sessions are the richest source of pitfalls. The prompt explicitly says:
|
||||
> "If the session failed or was partial, extract pitfalls and questions — these are the most valuable."
|
||||
|
||||
This reframes failure as valuable output, not noise to discard.
|
||||
|
||||
### The `evidence` field
|
||||
|
||||
Added to the original spec. Every extracted item must cite where in the transcript it came from. This:
|
||||
- Prevents hallucination (can't cite what isn't there)
|
||||
- Enables verification (reviewer can check the source)
|
||||
- Trains confidence calibration (the agent must find evidence, not just claim it)
|
||||
|
||||
### Token budget
|
||||
|
||||
Target: ~1,000 tokens for the prompt (excluding transcript).
|
||||
|
||||
```
|
||||
System prompt: ~50 tokens
|
||||
Rules: ~200 tokens
|
||||
Categories: ~150 tokens
|
||||
Confidence: ~100 tokens
|
||||
Output format: ~200 tokens
|
||||
Design notes: NOT included in prompt (documentation only)
|
||||
─────────────────────────────
|
||||
Total prompt: ~700 tokens
|
||||
```
|
||||
|
||||
Leaves ~300 tokens headroom for variable content (transcript insertion, edge cases).
|
||||
|
||||
### What this replaces
|
||||
|
||||
The v1 prompt had:
|
||||
- Verbose prose explanations (waste tokens for mimo)
|
||||
- No `evidence` field (hallucination risk)
|
||||
- No `meta` block (no session-level metadata)
|
||||
- No explicit handling of failed sessions
|
||||
- Example was too long (~150 tokens of example for a 1k prompt)
|
||||
|
||||
This v2 is tighter, more structured, and adds the evidence requirement that prevents the #1 failure mode of extraction prompts: generating plausible-sounding facts that aren't in the transcript.
|
||||
- Process the entire transcript, not just the beginning
|
||||
- Pay special attention to errors and corrections
|
||||
- Note any environment-specific details
|
||||
- Track tool-specific behaviors and quirks
|
||||
- If the session failed, focus on pitfalls and questions
|
||||
|
||||
Reference in New Issue
Block a user