Changelog

Fat traces, the run reviewer, and the new dashboard

• •ArgusRecorder: capture via LangChain callbacks, nothing patched. Keeps the dict each node returned — its update, not the merged state
• •Run reviewer: with node purposes, an LLM checker and verifier must agree before a heuristic finding fails CI; a finding no rule saw fails only when two models verify it
• •Whole-trace rules D1–D17: typo'd keys, vendor errors, unfollowed pagination, ungrounded numbers, near-miss IDs, claimed actions with no tool call, stuck loops
• •Dashboard: each step shows its tool calls and what the run reviewer verified; rerun from the failing step through your app factory
• •GitHub Action: uses: ArgusLabs-ai/ARGUS/.github/actions/argus-gate@v0.12.0
• •pytest --argus records through LangChain's configure hook — nothing patched — and binds each run to the test that produced it
• •argus ignore, argus check --strict warn_as_fail, and argus doctor checks for storage, keys and disk usage
• •consumers={"field": ["reader"]}: declare who reads what, so a field never written / written empty / dropped in between is blamed on the node responsible
• •argus ingest langsmith: grade a LangSmith export with no app and no graph. Skinny traces are refused, not graded green
• •report_tool_call(): file a tool the callback path cannot see, so a 404 body or a raised tool is not silently absent from the step
• •Parallel Send workers are siblings, not retries — a swallowed error in the first of five workers no longer hides behind the last one passing
• •A node's own verdict (status: denied, a linter's errors list) is a warning, not a CI fail. The same shape from a tool response stays critical
• •The judge reviews soft flags only: it cannot originate a fail, clear a hard fail, or move blame off an origin
• •Barren subgraphs: every inner node wrote something, all of it to inner-only keys, parent state unchanged (subgraph_no_contribution)
• •httpx traffic recorded via httpcore; a capture session that records nothing warns instead of writing an empty cassette
• •A ReAct agent's blank final reply (empty content, no tool calls) fails the gate; intermediate tool-calling turns stay exempt
• •A KeyError from a node's own conditional edge is blamed on that node, not on the upstream bystander that last touched the state
• •Consumer keys accept dotted paths (email.body) so a blanked nested leaf is caught; a top-level declaration still means the whole value
• •allow_empty reaches the writer's tool responses, so a clean sanctions screen returning hits: [] no longer fails CI. Undeclared empty retrieval stays critical

Findings, check JSON, replay patches

• •RunRecord.findings: one normalized failure list per run (older records back-filled on load)
• •argus check --format json and --fail-on; check grades the requested run
• •argus replay --set / --delete / --patch / --dry-run
• •pytest --argus watches ainvoke / stream / astream
• •chat_response and code_generation behavior profiles
• •Truncation, stale-context, and trailing etc. signatures; warning-severity matches no longer fail a node on their own
• •Dashboard /runs/<id> deep links

Double-encoded JSON detection

• •Rule 17: a string field that parses as a JSON object/array is flagged (json_in_string) — double-encoded JSON returned instead of a parsed structure
• •Advisory only (warning); fields where stringified payloads are expected (raw_response, log, payload, …) are skipped

argus show resolves the requested run

• •argus show <id> now renders the requested run instead of always showing the newest one
• •Unknown ids and ambiguous prefixes exit 1 with a clean message instead of a traceback

Nonzero exit on failed key commands

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

Silent no-op detection & per-invoke persistence

• •empty_output failure: a node returning a literal {} with successors downstream is flagged critical and blamed on the origin, not the crash site
• •An attached ArgusWatcher now persists a run record for every invoke() / stream() / batch(), not just the first
• •Wave E stability: retry/cycle false positives, validator severity, replay verdict, confidence explanation

GitHub & PyPI version alignment

• •Republish 0.9.2 + 0.9.3 as 0.10.1 so pip install argus-agents is no longer stuck on 0.10.0
• •argus check last / argus check <id> — exit 1 when a run was not clean
• •pytest --argus — auto-instrument LangGraph invokes in tests
• •argus init skills, terminal findings, and empty-dashboard path hint

CI Gate & pytest Plugin

• •argus check last / argus check <id> — exit 1 when a run was not clean (crash, silent failure, semantic fail)
• •pytest --argus — auto-instrument LangGraph invokes in tests; silent failures become test failures
• •Heuristics-only in pytest (judge off); tests that never invoke a graph are unchanged
• •Guide, README, and argus-debug skill updated for CI workflows

Detection Pipeline Hardening

• •Patched 4 detection pipeline bugs uncovered by 367-test audit suite
• •Crash sites demoted in degradation origin ranking — correlator now points at causes, not symptoms
• •Status rendering fixes across CLI and UI (VAR-91, VAR-92)
• •CI lint failures resolved across test suite

Decision Explanations & Tool Call Chains

• •Decision explanations for every detection — see why ARGUS flagged (or passed) each node
• •Tool call chain analysis across runs (VAR-10) — trace tool usage patterns over time
• •Graph edge routing avoids intermediate node overlap for cleaner pipeline visualizations
• •--json flag for argus show and argus list for scripted workflows

Evidence-Aware Judge & Visual Overhaul

• •Evidence-aware LLM judge with audit trail — evidence_considered and overridden_signals on every ruling (VAR-83, VAR-84)
• •Distinct node colors per failure type in execution graph — crashed (red), silent (amber), semantic (purple), degraded (orange), skipped (gray)
• •Root-cause misattribution fix when correlation engine and LLM investigation disagree
• •Differentiate skipped vs crashed nodes in conditional paths (VAR-61)
• •Latency degradation detection — flags nodes at 95%+ of timeout or suspiciously fast (VAR-8)

Production-Ready Configuration

• •Typed ArgusConfig with semantic judge failure policy (VAR-67, VAR-68, VAR-69)
• •Config parameter cross-validation at construction time (VAR-73)
• •Auto-detect cyclic graphs needing finalize() (VAR-70)
• •Pattern-based secret detection + custom redaction functions (VAR-72)
• •Run persistence sampling and schema versioning (VAR-71)
• •finalize() idempotency, concurrent safety, and dry-run mode (VAR-75)
• •Lazy imports for optional dependencies (VAR-74)
• •Ctrl+C graceful shutdown, auto-finalize, guide overhaul

Semantic Intelligence

• •Embedding-based semantic similarity matching in heuristic engine
• •Auto-generalize LLM-suggested signatures before storing
• •Signature effectiveness tracking, auto-prune dead signatures, and UI overhaul
• •Confidence-gated LLM escalation at heuristic + correlator layers
• •Auto-locate source files for replay via LLM + AST analysis
• •Input-output coherence checks (VAR-7)
• •Loop-aware inspection + mandatory LLM loop analysis (VAR-62)
• •Reducer-aware inspection — suppress false-positive selective_attention_reduction for LangGraph reducer fields
• •Discord webhook integration with rich diagnostic embeds

Adaptive Learning & LLM Judge

• •Shared signature library with approvals page for community-contributed patterns
• •LLM semantic judge as final authority with override controls
• •Feedback loop — disputed detections refine future heuristics
• •Auto LLM comparison for replay runs with per-node smart diff
• •Diagnostic report system — send run reports from the dashboard
• •Interactive node inspector in the UI
• •All LLM calls routed through ARGUS proxy — no user API key required for judge features

LLM Semantic Judge

• •Optional LLM semantic judge for subtle quality issues beyond pattern matching
• •Dark theme for the web dashboard

Integration & Rerun Reliability

• •watch_compiled() — attach ARGUS to already-compiled graphs, no more "must call before compile()" trap
• •Reducer-aware state merging — list fields (messages, tool calls) now append correctly during reruns instead of being overwritten
• •HTTP recording — opt-in record_http=True captures all API calls for fully deterministic reruns
• •argus doctor — 5-second diagnostic command checks Python, LangGraph, storage, and rerun readiness
• •Renamed "replay" to "rerun" across all UI, CLI, and docs — honest about what it actually does
• •Live-call warning in CLI — explicitly tells users that external API calls execute live unless HTTP recording is enabled

Detection Engine Upgrade

• •7 new failure patterns: truncated outputs, retrieval quality, hallucinated success, confidence-behavior mismatch, shallow summaries, partial tool failures, corrupted structured output
• •Single-node rerun — re-run one node in isolation without rebuilding the full graph
• •Inline before/after diff shown in CLI immediately after rerun completes
• •AI investigation now wired into rerun results via stored node refs
• •14 new semantic signatures: hedging phrases, placeholder variants, corrupted markers
• •Short-string repetition detection fixed (N/A × 4 and similar patterns now caught)
• •Rerun tree visible in run detail Overview tab

Master-Detail UI & Factory-Free Rerun

• •Factory-free rerun — node function refs stored per-run, no --app flag required
• •Master-detail layout: resizable split panel, runs list left, full detail right
• •Overview tab: status card, pipeline overview, metrics grid, AI analysis summary, execution timeline, rerun branches
• •Pipeline tab: node-by-node inspector with expandable input/output per step
• •AI Analysis tab: LLM root cause investigation panel with confidence score
• •Correlations, State, and Logs tabs added

Beta Testing Rollout

• •LLM token/cost tracking — auto-extracts usage from node outputs, shows cost per node and total
• •Redesigned sidebar into Observe / Analyze sections
• •Rerun from subdirectories — runs recorded from child folders are now found and rerunnable
• •Changelog page with full version timeline
• •Report the Dev page for bug reports and feature requests
• •Restored and expanded test suite

Rerun wired to UI

• •Rerun from UI — hover any step, click "rerun from here", no CLI needed
• •App factory input persists to .argus/config.json automatically
• •Auto-compare after rerun — navigates to diff view showing original vs rerun
• •Eval metrics panel in compare page (failure count, severity, success rate)
• •--app flag on argus ui for startup config

Auto-login & interrupt stitching

• •Auto-login in argus ui — reads CLI credentials, no second OAuth flow
• •Interrupted runs + resume continuations stitched into a single merged view
• •Resume runs hidden from top-level list to avoid duplicates

Cloud storage

• •Google login and Supabase cloud storage
• •Background sync of runs to cloud

Web Dashboard

• •argus ui — web dashboard served by a pure-Python HTTP server, no Node.js required
• •Runs list with status, duration, pass rate stats
• •Per-run detail view with step-by-step inspection
• •Side-by-side run comparison at /compare
• •Zero extra dependencies — just Python

Silent failure enhancements

• •Strict mode for inspector
• •Nested error scan in tool outputs
• •Generic list type checking

Parallel Execution

• •Parallel nodes grouped in a parallel panel in argus show
• •Graph topology diagram above the step list
• •Silent failure detection is parallel-aware — only blames nodes for fields they wrote
• •Root cause chain excludes innocent parallel siblings

Run Differentiator

• •argus diff — compare any two runs node-by-node
• •Status changes with FIXED / REGRESSION labels
• •Inspection diff, validator result flips, output field deltas
• •Frozen node and duration change tracking

Tool Call Failure Detection

• •Detects error_response, rate_limit (HTTP 429), and empty_result from tool calls
• •Tool failures show inline under the node that swallowed them

Deterministic Replay

• •Replay plays back exact outputs from the original run — no live LLM calls
• •Reproducible replays with zero configuration

MVP

• •Core monitoring: ArgusWatcher for LangGraph, ArgusSession for any framework
• •Silent failure detection: missing fields, type mismatches, empty outputs
• •Semantic signature registry for placeholder/degraded LLM outputs
• •Root cause analysis chain
• •CLI: argus show, argus replay
• •Local storage in .argus/runs/