Validated at every boundary.
ExecutionAuditLog, ExtractedHeuristic and DocUpdatePayload as Pydantic models and JSON Schemas.
The memory core of the autonomous AI software factory
The ledger keeps the scars.
The rules keep the lesson.
The Chronicler watches the EdgeRecord ledger, extracts failure heuristics through local Ollama, and commits hardened rules and lore docs back to Git — headless, local-first, with a path-traversal guard before every write.
chronicler doctor: is the local Ollama reachable, is the ledger readable, and is the target repository a Git repository?01 / Quick start
Point it at a ledger and a repository.
No cloud inference. No paid fallback. The model is the one you run.
# install pip install the-chronicler # point it at your local ledger and target repo export CHRONICLER_LEDGER_PATH=~/.taskrivet/ledger.sqlite export CHRONICLER_REPO_ROOT=~/Projects/taskrivet # sanity-check the black ops setup chronicler doctor # ingest one batch and commit what it learns chronicler run --once
02 / What it does
A cycle, with receipts.
Every cycle reports what it read, what it learned and what it committed.
No failure is ever skipped.
The ledger is read oldest-first, so the watermark advances in order and a backlog drains completely instead of hiding behind the newest row.
One bad scar cannot stop the batch.
A malformed local-LLM reply is reported in failed_audits while the healthy failures behind it still land.
Stable ids, stable dedupe.
Heuristic ids are derived from the failure category and fingerprint, so the same lesson is not learned twice.
A path guard before every write.
The committer refuses an unsafe path. Push is configurable, and each cycle captures its own commit receipts.
Your model, your machine.
Analysis routes through a local Ollama endpoint. Nothing is sent to a proprietary cloud LLM.