Reflex

Reflex is a local, evidence-first observability and learning loop for agent work. It ingests safe projections of sessions, tool calls, CI results and delivery receipts from the Subactor, Semcod and Autogrammar ecosystem; it then normalizes repeated events, detects recurring failure patterns and emits bounded improvement proposals.

What is delivered

The first release provides a deterministic, stdlib-only pipeline for JSONL, NDJSON and bounded text evidence:

• secret and home-directory redaction, event normalization and duplicate collapse;
• recurring-pattern detection for governance, dependency, tool-contract, credential-readiness, remote-rate-limit and test-feedback failures;
• an optional SQLite store containing normalized evidence only, so observations can accumulate across sessions without retaining raw transcripts;
• a closed reflexdsl request format for bounded INGEST, ANALYZE, REPORT, STATUS and PROPOSE operations;
• an offline status projection with issue creation lists, open-resource counts, closed/merged pull requests, last closure time and closure history;
• an optional SubLLM reflection step that produces advisory hypotheses only.

The analysis and ownership recommendations are recorded in docs/analysis/reflex-learning-loop.md.

Quick start

From the assembled checkout:

python -m venv .venv
pip install -e .
reflex analyze --source /path/to/safe/logs --store .subactor/cache/reflex.sqlite
reflex analyze --source /path/to/safe/logs --llm
reflex status --source /path/to/github-events.jsonl --format text

For a bounded DSL request:

reflex dsl --request '```reflexdsl
ANALYZE ROOT "/path/to/safe/logs" MIN_REPEAT 3
```' --allowed-root /path/to/safe/logs

For lifecycle status, use the closed DSL command:

reflex dsl --request '```reflexdsl
STATUS ROOT "/path/to/github-events.jsonl" FORMAT text
```' --allowed-root /path/to

Use a path outside the repository, or the ignored .subactor/cache area, for the SQLite store. Reflex never treats a proposal as permission to create a ticket, change code, merge a branch, deploy a service or publish a release.

Operating boundary

Only bounded projections should be supplied as input. Raw chat transcripts, credentials, private keys and unrestricted environment dumps are not inputs to this project. Source lines are represented by digests and short redacted summaries. The LLM is downstream of deterministic normalization and its output is advisory; protected governance, ticket ownership and delivery controllers remain authoritative.

The repository is intentionally initialized with the Wellmanifest governance contract before implementation. See TODO.md for the next bounded integration steps and docs/README.md for the analysis index.