[0.8.0] — 2026-07-13

The observability release: the harness now covers the agents you use as well as
the agents you build. New positioning: pytest + observability infrastructure for
AI agents.

Added

• proof watch: live coding-agent observability. Attaches to the coding agent
  working in your repo (Claude Code and Cursor natively, anything else via the git
  working tree) and screens the session as it happens. Two cadences, both in seconds:
  --screen-every (default 30) runs the intelligent risk screening at 0 tokens;
  --interval (default 120) runs the harness evaluation + dashboard upsert. The
  growing session is upserted as a SINGLE live run (stable session_key), so the
  trajectory fills in near real time. Observe only; never gates the agent. --once
  takes a single snapshot for CI; Ctrl-C flushes a final live=false snapshot.
• Canonical intent trajectory (harness synthesis). One harness LLM call reads the
  whole session and labels each turn with a standardized <verb> <object> intent
  (resolved from full session context, not a paraphrase of the prompt), what the agent
  did, and a risk note. Synthesis is signal driven (new risk, a batch of new intents,
  or a periodic cap); --analyze-every-interval forces it on every interval with new
  turns. Narration is sticky: a turn labeled once never downgrades on later scans.
• Preflight checks. proof watch verifies the tool connection, probes the harness
  LLM (probe_llm: one minimal call), and prints the upload target before the loop
  starts, so a bad key or model surfaces up front instead of silently degrading.
• Honest narration accounting. Eval tokens accumulate across scans
  (narration_tokens), the narrated flag is true only when the LLM actually spent
  tokens, and a failed synthesis surfaces narration_error plus a terminal warning
  (missing key, bad model, rate limit) instead of silently showing prompt-like intents.
• Cursor adapter. Reads Cursor's per-workspace session store; --tool auto falls
  back Claude Code, then Cursor, then the git diff.
• Idempotent proof session: a session carries a stable session_key (from the
  transcript), so re-running it updates the same dashboard run instead of duplicating it.

Changed

• --interval is now seconds (was minutes), the same unit as --screen-every;
  session metadata carries interval_seconds and the live scan state (live,
  last_scan, next_scan) via build_session_payload(..., session_key=, watch=).
• Secret redaction hardened. Key shaped tokens and emails are masked in every
  uploaded free text field: prompts, intent previews, event messages and targets, and
  access map commands. Finding evidence now carries the event timestamp (ts) so the
  dashboard can attribute findings to trajectory turns.
• README repositioned around the two planes (evaluate the agents you build,
  observe the agents you use) with a new "Observe the coding agents you use" section
  and CLI reference entries for proof watch / proof session. The ProofAgent
  Governance platform is documented as a separate commercial product; --upload
  stays the single optional integration point.
• PROOFAGENT_API_BASE_URL documented as the supported way to repoint --upload
  (on-prem / Enterprise / local stacks); module docs previously claimed the
  destination could not be repointed.

Fixed

• proof watch cumulative token bookkeeping no longer overwrites the deep assessment
  spend in total_tokens when --assess grades a slice in the same scan.
• Offline intent labels (actionize) strip conversational preamble (typo tolerant),
  so a turn is labeled by its request, not its throat clearing.
• Repo hygiene for OSS: removed stray tracked artifacts (mt.json, mt.md, 0),
  ignored /.claude/ so local editor state can never ship in an sdist.