Previous version was outdated (said scripts were 'not implemented').
Updated to reflect actual state: 18 scripts, 14 test files, populated
knowledge store, active development.
2026-04-21 04:43:54 +00:00
3 changed files with 199 additions and 186 deletions
**What:** A system that turns 1B+ daily agent tokens into durable, compounding fleet intelligence.
Every agent session starts at zero. The same HTTP 405 gets rediscovered as a branch protection issue. The same token path gets searched from scratch. Intelligence evaporates when the session ends.
**Why:** Every agent session starts at zero. The same mistakes get made repeatedly — the same HTTP 405 is rediscovered as a branch protection issue, the same token path is searched for from scratch. Intelligence evaporates when the session ends.
**How:** Three pipelines form a compounding loop:
Compounding-intelligence solves this with three pipelines forming a loop:
```
SESSION ENDS → HARVESTER → KNOWLEDGE STORE → BOOTSTRAPPER → NEW SESSION STARTS SMARTER
@@ -18,222 +18,234 @@ SESSION ENDS → HARVESTER → KNOWLEDGE STORE → BOOTSTRAPPER → NEW SESSION
MEASURER → Prove it's working
```
**Status:**Early stage. Template and test scaffolding exist. Core pipeline scripts (harvester.py, bootstrapper.py, measurer.py, session_reader.py) are planned but not yet implemented. The knowledge extraction prompt is complete and validated.
---
**Status:**Active development. Core pipelines implemented. 20+ scripts, 14 test files, knowledge store populated with realdata.
| `gitea_issue_parser.py` | Parse Gitea issues for knowledge extraction |
### Pipeline 2: Bootstrapper
### Automation
**Status:** Not implemented.
| Script | Purpose |
|--------|---------|
| `session_pair_harvester.py` | Extract training pairs from sessions |
Queries knowledge store before session start. Assembles a compact 2k-token context from relevant facts. Injects into session startup so the agent begins with full situational awareness.
### Pipeline 3: Measurer
**Status:** Not implemented.
Tracks compounding metrics: knowledge velocity (facts/day), error reduction (%), hit rate (knowledge used / knowledge available), task completion improvement.
The 2k-token injection that a new session receives. Assembled from the most relevant knowledge items for the current task, filtered by confidence > 0.7, deduplicated, and compressed.
### Knowledge Index (knowledge/index.json)
Machine-readable fact store. 12KB, populated with real data. Categories: fact, pitfall, pattern, tool-quirk, question.
---
## Knowledge Store
```
knowledge/
├── index.json # Master fact store (12KB, populated)
1.**Harvester integration test** — Does the prompt actually extract correct knowledge from real transcripts?
2.**Bootstrapper test** — Does it assemble relevant context correctly?
3.**Knowledge store test** — Does the index.json maintain consistency?
4.**Confidence calibration test** — Do high-confidence facts actually prove true in later sessions?
5.**Deduplication test** — Are duplicate facts across sessions handled?
6.**Staleness test** — How does the system handle outdated knowledge?
---
1.**session_reader.py** — No dedicated test file (tested indirectly)
2.**sampler.py** — No test file (scoring logic untested)
3.**session_metadata.py** — No test file
4.**validate_knowledge.py** — No test file
5.**knowledge_staleness_check.py** — Tested but limited
## Security Considerations
1.**No secrets in knowledge store** — The harvester must filter out API keys, tokens, and credentials from extracted facts. The prompt constraints mention this but there is no automated guard.
### API Key Handling
-`harvester.py` reads API key from `~/.hermes/auth.json` or env vars
- Key passed to LLM API in request headers only
- No key logging
2.**Knowledge poisoning** — A malicious or corrupted session could inject false facts. Confidence scoring partially mitigates this, but there is no verification step.
### Knowledge Integrity
-`validate_fact()` checks schema before writing
-`deduplicate()` prevents duplicate entries via fingerprint
-`knowledge_staleness_check.py` detects when source code changed but knowledge didn't
- Confidence scores prevent low-quality knowledge from polluting the store
3.**Access control** — The knowledge store has no access control. Any process that can read the directory can read all facts. In a multi-tenant setup, this is a concern.
### File Safety
- Knowledge writes are append-only (never deletes)
- Bootstrap context is truncated to budget (no prompt injection via knowledge)
4.**Transcript privacy** — Session transcripts may contain user data. The harvester must not extract personally identifiable information into the knowledge store.
parser.add_argument("--repo",default=".",help="Path to repository to analyze")
parser.add_argument("--json",action="store_true",help="Output as JSON")
@@ -521,6 +521,7 @@ def main():
help="Slow test threshold in seconds")
args=parser.parse_args()
globalSLOW_TEST_THRESHOLD_S
SLOW_TEST_THRESHOLD_S=args.threshold
ifnotos.path.isdir(args.repo):
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
