Work / Applied AI engineering
Experimental projectJev Context Pruner CLI
Testing model-assisted conversation pruning with deterministic safeguards and evidence evaluation.
A Python CLI experiment that uses typed model judgments inside a deterministic policy to keep, truncate, or drop eligible conversation tool pairs. It includes protected inputs, full-original fallback, traces, and evidence-based evaluation.
The question behind the tool
Could a model help reduce a saved development conversation while a separate, fixed policy keeps the decision rules reviewable? This project explores that question as a standalone file-conversion and evaluation CLI.
It does not connect to or replace the live conversation history in Codex, and it never executes tools or commands found in a transcript.
A constrained decision path
The CLI validates a transcript snapshot, protects messages and sensitive or uncertain tool pairs, then asks Jev two typed questions about each remaining eligible pair. Deterministic thresholds choose KEEP, TRUNCATE, or DROP; the model does not rewrite the conversation freely.
- Messages, explicit pins, writes, unknown effects, errors, and the most recent pairs are kept in full without model questions.
- DROP removes a tool call and its result together; TRUNCATE can only shorten result content.
- Missing credentials, provider errors, invalid responses, or request limits return the complete original case.
- Live requests send the task and transcript events to the provider; the tool does not automatically remove identifying information. Use only authorized public, synthetic, or personal data.
Evaluation that preserves failures
The repository includes 12 synthetic evaluation cases, with separate gold evidence and fixtures. The fixed fixture run evaluated 11 cases, recorded one no-op and no fallbacks, retained 12 of 13 evidence items overall and 5 of 6 eligible items, and reduced canonical characters by about 13.0%.
A live run recorded on 22 September 2026 evaluated 11 cases, with one no-op and no fallbacks. It retained 12 of 13 evidence items overall and 5 of 6 eligible items, with about 18.4% fewer canonical characters. Its lost item was “START ok” in case03; the fixture run lost a different item.
What these results do and do not show
The small synthetic suite shows how the evaluator can expose a lost evidence item. It does not establish broad accuracy or that Jev consistently outperforms a baseline. Character reduction is not a token, cost, or runtime-context measurement, and one live run is not a general performance claim.
The project is experimental. It has no Codex adapter, hook, interface, database, agent loop, tokenizer, or automatic tuning. Its public design and records are available in the linked repository.