Flow Indicator

flow-indicator

[pate.sh](https://pate.sh) | [provenance](https://tgp.ski) | [commands](docs/COMMANDS.md) | [herdr](docs/HERDR.md) | [integrations](docs/INTEGRATIONS.md) | [profiles](docs/PROFILES.md) | [inference](docs/INFERENCE.md) | [calibration](docs/CALIBRATION.md) | [events](docs/EVENTS.md) | [metrics](docs/METRICS.md)

A local meter for when cheap agent steering gives way to repeated correction, constraint restatement, repair, and operator serialization.

It answers one question about an interaction stream:

Is the operator advancing the work, or spending increasing effort keeping the agent aligned with work already established?

Stdlib-only Go, one binary.

What it reads

An ordered stream of interaction records, through a harness:

harness	input
claude-code	a Claude Code session file, ~/.claude/projects//.jsonl
codex	a Codex rollout, ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl
opencode	a session in opencode's SQLite store, #
qwen	a Qwen Code session file, ~/.qwen/projects//chats/.jsonl
generic	JSONL with at least stream_id, seq, speaker, text

Each harness owns every field name of its own format. No other package reads them. Replay and the live meter take the same names, so anything that can be replayed can be watched.

What it writes

Per stream, under `$XDG_DATA_HOME/flow-indicator/sessions//`:

file	role
observations.jsonl	append-only, observed events
classifications.jsonl	append-only, classifier candidates
obligations.jsonl	append-only, obligation transitions
repairs.jsonl	append-only, repair episode transitions
metrics.jsonl	append-only, derived metrics, regime and epoch changes
source.json	source path, byte offset, record count
summary.json, metrics.json, report.md, timeline.csv	projections, rebuilt from the events above

Th