v0.11.0

• RunRecord.findings — one Finding per inspection, validator, anomaly, judge, crash, and tool-chain signal (stable id, full-sentence reason, source). Older records are back-filled on load. (#60)
• argus check --format json and --fail-on crashed,silent_failure,interrupted (#61)
• argus check grades the requested run, not the newest. (#64)
• argus replay accepts --set, --delete, --patch, and --dry-run.
• pytest --argus watches ainvoke / stream / astream. (#66)
• CI runs pytest tests/test_argus_ci_gate.py --argus on every PR. (#68)
• Behavior profiles chat_response and code_generation. (#65)
• Truncation, stale-context, and trailing etc. signatures. (#43, #45)
• Warning-severity semantic signals no longer fail a node on their own. (#47)
• Dashboard /runs/<id> deep links. (#63)

pip install argus-agents==0.11.0

v0.10.5

Rule 17 — double-encoded JSON detection: a string field that parses as a JSON object/array is flagged json_in_string (warning, advisory only). Fields where stringified payloads are expected (raw_response, log, payload, …) are skipped. (#26)

v0.10.4

argus show <id> now resolves the requested run instead of always rendering the newest one; unknown ids and ambiguous prefixes exit 1 with a clean message instead of a traceback. (#33, #39)

v0.10.3

argus key set / argus key use now exit 1 on failure (unknown provider, no saved key) so scripts gating BYOK setup can detect it. (#40)

v0.10.2

Wave E stability fixes, silent no-op detection, and per-invoke persistence.

• Wave E: retry/cycle false positives, validator severity, replay verdict, confidence explanation (#32)
• empty_output failure: a node returning a literal {} with successors downstream is flagged critical and blamed on the origin, not the downstream crash site (#31). Only literal {} is flagged; dicts with keys (even empty-valued) and router/conditional nodes are exempt
• Per-invoke persistence: an attached ArgusWatcher now writes a separate run record for every invoke() / stream() / batch(), not just the first
• Restore the test_no_ordering_anomaly_correct_order case that had been shadowed as dead code; single-source the correlator confidence weights and _INACTIVE_STATUSES constant

Full changelog: https://github.com/ArgusLabs-ai/ARGUS/blob/master/CHANGELOG.md

v0.10.1 — GitHub & PyPI version alignment

Why this release

PyPI listed 0.10.0 as latest (0.10.0 > 0.9.3), while GitHub Latest was v0.9.3. 0.10.0 shipped on 14 Aug (same era as 0.9.1), so pip install argus-agents skipped the 0.9.2 / 0.9.3 work.

This tag republishes that work as 0.10.1 so GitHub and PyPI advertise the same package.

What's included (from 0.9.2 + 0.9.3)

• argus check last / argus check <id> — CI gate; exit 1 on crash, silent failure, semantic fail, missing fields, or tool failures
• pytest --argus — auto-wrap LangGraph compile() / invoke() in tests; unclean runs fail the test
• argus init writes Cursor/Claude skills that attach ArgusWatcher and debug from .argus/runs
• Failed runs print a short [argus] finding; clean runs stay silent
• Empty dashboard shows which .argus path is served

Install

pip install argus-agents==0.10.1

After the publish workflow finishes, pip install -U argus-agents also gets this version.

v0.9.3 — CI Gate & pytest Plugin

What's new

• argus check last / argus check <id> — CI gate that exits 1 on crash, silent failure, semantic fail, missing fields, or tool failures
• pytest --argus — auto-wraps LangGraph compile() / invoke() in tests; unclean runs fail the test without needing ArgusWatcher in the test file
• Heuristics-only in pytest (judge off); tests that never invoke a graph are unchanged

Docs

• Landing page guide updated with CI & testing section
• Website changelog updated
• argus-debug skill updated (refreshed on next argus init --force)
• README already documented in prior commits

Install

pip install argus-agents

v0.9.2

• argus init writes Cursor/Claude skills that attach ArgusWatcher and debug from .argus/runs
• Failed runs print a short [argus] finding; clean runs stay silent
• Empty dashboard shows which .argus path is served

v0.10.0 — One install, one attach()

What's new

Install

• pip install argus-agents is the full product (CLI + LangGraph + UI). No [cli] / [langgraph] extras needed.

Setup

• One API: ArgusWatcher().attach(graph) works on a StateGraph or a compiled app.
• invoke() / ainvoke() persist automatically. finalize() is optional.
• batch / abatch / stream save once when the outermost call returns (not after the first item).

LLM judge

• Opt-in with argus key set (or OPENAI_API_KEY / Anthropic / Google).
• No key → heuristic-only. No argus-agents[llm] extra.
• argus doctor reports BYOK vs heuristic-only.

Docs

• README, dashboard guide, and arguslabs.in match this path.

v0.9.1

• pip install argus-agents ships CLI, LangGraph adapter, and UI
• ArgusWatcher.attach(graph) is the one API (StateGraph or compiled); invoke() / batch() / stream() persist — no finalize() required
• LLM judge is opt-in via argus key set (or env); no [llm] extra. No key → heuristics only
• README, guide, and arguslabs.in docs match install + attach