v0.32.0 — Self-Contained HTML Session Replay Viewer

Self-Contained HTML Session Replay Viewer

agent-strace replay --format html generates a single-file HTML viewer for any recorded session. No server, no dependencies — open it in any browser.

agent-strace replay --format html
agent-strace replay --format html --output review.html SESSION_ID

Viewer features:

• Animated event timeline with configurable playback speed (up to 4×)
• Scrubber bar for jumping to any point in the session
• Running cost counter updated as events play
• Click-to-expand event detail (full JSON payload)
• Color-coded event types: tool calls, LLM requests, file ops, errors
• Pause/resume and show-all controls
• Dark theme, zero external dependencies (no CDN, no fonts)

All event data is embedded as a JSON constant in the HTML file. Useful for sharing sessions with teammates or attaching to PR reviews without requiring them to install anything.

v0.31.0 — Agent Standup Report from Session Trace

Agent Standup Report from Session Trace

agent-strace standup generates a structured standup report from a session trace — no LLM call required.

agent-strace standup
agent-strace standup --session SESSION_ID

Report sections:

What the agent did:

• Files read and modified
• Approaches tried, including abandoned ones (detected from retry patterns)
• New dependencies added (npm install, pip install, etc.)

What it was uncertain about:

• TODO / FIXME / assumption comments written into files

What to review carefully:

• Large changes (>100 lines), new dependencies, auth and migration patterns

Stats: tool calls, context resets, retries, errors

Useful for async teams where the agent runs overnight and a human needs a quick brief before picking up the work.

v0.30.0 — On-Call Readiness Report for Agent-Modified Files

On-Call Readiness Report for Agent-Modified Files

When an agent has been writing code, the human on call may not have read it. agent-strace oncall cross-references agent-modified files from the trace store against git history to surface cognitive gaps before a rotation.

agent-strace oncall --rotation-start 2026-04-25
agent-strace oncall --rotation-start 2026-04-25 --scope "src/payments/**"

For each file the agent has written in the last N days, the report shows:

• How long ago it was modified
• How many lines changed (from git log --numstat)
• Estimated reading time (~200 lines/minute)
• Total catch-up time before rotation

--scope filters to a file glob. --since-days controls how far back to scan sessions (default: 30).

v0.29.0 — Context Freshness Check Before Starting a Session

Context Freshness Check Before Starting a Session

Before handing a task to an agent, it helps to know how stale its last view of the codebase is. agent-strace freshness compares the current state against what the agent last saw, using git diff between the last session timestamp and HEAD.

agent-strace freshness
agent-strace freshness --since 2026-04-01 --scope "src/**"

Report includes:

• Files changed since the last session (or since --since date)
• Per-file change type (modified / added / deleted / renamed) and line count
• Freshness score 0–100 (100 = nothing changed since last session)
• Estimated catch-up reading time for in-scope files

Scope is auto-detected from CLAUDE.md / AGENTS.md scope sections, or overridden with --scope. No API calls required.

v0.28.0 — A2A Protocol Support with Cross-Agent Trace Correlation

A2A Protocol Support with Cross-Agent Trace Correlation

First-class support for agent-to-agent calls following the Google A2A spec. A2A calls are captured as TOOL_CALL events with event_subtype=a2a_call, so they are backward-compatible with all existing replay and export tooling.

agent-strace a2a-tree
agent-strace a2a-tree SESSION_ID --format json

New capabilities:

• Detects A2A calls by path, header, and body heuristics
• Builds the full agent call graph by following sub_session_id links and parent_session_id back-references
• Renders the call graph as an ASCII tree
• Exports the graph as OTLP-compatible spans for Jaeger, Tempo, or any OpenTelemetry backend

Child sessions are linked via parent_session_id and parent_event_id in session metadata.

v0.27.0 — Token Inflation Calculator Across Model Versions

Token Inflation Calculator Across Model Versions

Switching model versions can silently increase your token bill. agent-strace inflation measures the tokenizer cost impact before you commit to an upgrade — no API calls required.

agent-strace inflation
agent-strace inflation --compare claude-opus-4-6,claude-opus-4-7 --sessions 30

Applies per-model inflation factors to stored session content and breaks down the impact by content type: system prompt, tool definitions, user messages, assistant messages. Projects per-session, daily, and monthly cost delta.

Known inflation factors (relative to claude-opus-4-6 baseline):

Model	Factor
claude-opus-4-7	1.38× (community median: 1.3–1.47×, April 2026)
gpt-4o	1.05× (cl100k_base → o200k_base)

Prefix matching handles versioned names (e.g. claude-opus-4-7-20260101).

v0.26.0 — Personal Agent Cost-Efficiency Curve

Personal Agent Cost-Efficiency Curve

agent-strace curve analyses your stored session history and tells you which task types are worth delegating to an agent — and which ones cost more than they save.

agent-strace curve
agent-strace curve --min-sessions 10 --export csv

Sessions are classified into 10 task types (unit tests, debugging, refactoring, architecture, etc.) by keyword matching on agent name and command. Each type is compared against a community sweet-spot benchmark to produce a per-type verdict: efficient / over sweet spot / do this yourself.

Potential monthly savings are calculated for task types running above 1.5× their sweet spot. --export csv outputs raw data for external analysis.

v0.25.0 — Shadow AI Detection for Repositories

Shadow AI Detection for Repositories

agent-strace audit-tools scans a git repository for AI tool usage signatures — no network calls, no API keys required.

agent-strace audit-tools
agent-strace audit-tools --repo . --since "90 days ago" --approved cursor,copilot

Detected tools: Claude Code, Cursor, GitHub Copilot, Codex/ChatGPT, Windsurf, Aider — identified via file signals (.cursorrules, CLAUDE.md, .github/copilot-instructions.md, etc.) and commit message patterns.

What it flags:

• Unapproved tools against a configurable --approved list
• Unknown LLM API endpoints found in .env git history
• PII patterns in recently committed files

Useful for security audits and compliance reviews where you need to know which AI tools have touched a codebase.

v0.24.0 — Rich Side-by-Side Session Comparison

Rich Side-by-Side Session Comparison

agent-strace diff now produces a structured comparison table across cost, duration, tool calls, redundant reads, context resets, files modified, and errors — with a deterministic verdict that requires no LLM.

agent-strace diff SESSION_A SESSION_B --compare

New metrics:

• Redundant reads — files read more than once in a session
• Context resets — LLM requests separated by a >120s gap
• Approach divergence — first phase pairs where behaviour differs
• Verdict — derived from cost, duration, and error comparisons

Useful for regression testing: run the same prompt twice (or against two model versions) and get a structured diff you can assert on in CI.

v0.23.0 — Rule-Based Kill Switch for Runaway Sessions

Rule-Based Kill Switch for Runaway Sessions

Agents can go off the rails — burning tokens, modifying files they should not touch, or looping on failing tests. This release adds a declarative rules engine to agent-strace watch that can pause, kill, or alert when a session crosses a threshold.

agent-strace watch --rules .watch-rules.json
agent-strace watch --rules .watch-rules.json --dry-run

Rule conditions:

• files_modified — number of distinct paths written
• cost_usd — estimated spend so far
• consecutive_test_failures — repeated test failures in a row
• duration_minutes — wall-clock session length
• file_path — glob match on any file touched

Actions:

• pause — SIGSTOP the agent process (resume with SIGCONT)
• kill — SIGTERM, then SIGKILL after 5s; auto-generates a postmortem
• alert — log only, no interruption

--dry-run evaluates rules without acting — useful for tuning thresholds before going live.