agent-log-verifier 0.3.0__tar.gz

agent_log_verifier-0.3.0/.github/workflows/ci.yml

@@ -0,0 +1,33 @@
+name: CI
+
+# Mirrors scripts/gate.ps1 (the local Definition-of-Done gate):
+# pytest, mypy --strict, ruff. Runs on Ubuntu and Windows because path
+# normalization across POSIX/Windows forms is core domain logic here,
+# not an incidental portability concern.
+
+on:
+  push:
+    branches: [master]
+  pull_request:
+
+jobs:
+  gate:
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, windows-latest]
+        python-version: ["3.12", "3.13"]
+    runs-on: ${{ matrix.os }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+      - name: Install
+        run: python -m pip install -e .[dev]
+      - name: pytest
+        run: python -m pytest -q
+      - name: mypy --strict
+        run: python -m mypy --strict src
+      - name: ruff
+        run: python -m ruff check src tests

agent_log_verifier-0.3.0/.github/workflows/release.yml

@@ -0,0 +1,47 @@
+name: Release to PyPI
+
+# Publishes on version tags (vX.Y.Z) via PyPI Trusted Publishing
+# (OIDC) -- no API token is stored anywhere. One-time setup on
+# pypi.org: add this repository as a (pending) trusted publisher for
+# the `agent-log-verifier` project with workflow file
+# `release.yml` and environment `pypi`.
+
+on:
+  push:
+    tags: ["v*"]
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Gate before building
+        run: |
+          python -m pip install -e .[dev]
+          python -m pytest -q
+          python -m mypy --strict src
+          python -m ruff check src tests
+      - name: Build sdist and wheel
+        run: |
+          python -m pip install build
+          python -m build
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist
+          path: dist/
+
+  publish:
+    needs: build
+    runs-on: ubuntu-latest
+    environment: pypi
+    permissions:
+      id-token: write
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: dist
+          path: dist/
+      - uses: pypa/gh-action-pypi-publish@release/v1

agent_log_verifier-0.3.0/.gitignore

@@ -0,0 +1,58 @@
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.egg-info/
+.eggs/
+build/
+dist/
+.pytest_cache/
+.coverage
+.coverage.*
+htmlcov/
+.mypy_cache/
+.ruff_cache/
+
+# venv
+.venv/
+venv/
+env/
+
+# Editors
+.vscode/
+.idea/
+
+# OS
+.DS_Store
+Thumbs.db
+
+# Real logs / secrets — this tool reads Claude Code transcripts which may
+# contain confidential business data. Never commit real logs, only the
+# synthetic fixtures under tests/fixtures/synthetic_logs/ and
+# tests/fixtures/project_layout/ (both are 100% synthetic, hand-authored
+# JSONL/meta.json used by the test suite -- never copies of real logs).
+*.jsonl
+!tests/fixtures/synthetic_logs/**/*.jsonl
+!tests/fixtures/project_layout/**/*.jsonl
+*.meta.json
+!tests/fixtures/synthetic_logs/**/*.meta.json
+!tests/fixtures/project_layout/**/*.meta.json
+.agent-log-verifier/
+config.toml
+.env
+.env.*
+
+# scripts/gen_demo_logs.py output: regenerated on demand (README
+# quickstart step), never committed. Keeping this out of git means the
+# *.jsonl allowlist rules above never need a demo/-specific exception.
+/demo/
+
+# `alv dashboard` output: a self-contained HTML report generated from
+# transcript scan results. Finding.message content (including
+# confidential-looking paths) is deliberately rendered as-is for local
+# viewing (see report/html_output.py's docstring) rather than redacted,
+# so the generated file itself must never be committed -- only the
+# renderer code that produces it. Matches the CLI's default output
+# filename anywhere in the tree (an explicit --output path outside the
+# repo is the operator's own responsibility).
+alv-dashboard.html

agent_log_verifier-0.3.0/CHANGELOG.md

@@ -0,0 +1,168 @@
+# Changelog
+
+All notable changes to agent-log-verifier are documented here.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [Unreleased]
+
+## [0.3.0] - 2026-07-14
+
+First public release (GitHub + PyPI).
+
+### Changed
+- Demo generator timestamps are generation-relative (anchor: 90
+  minutes before run time; the stopped-subagent scenario sits 45
+  minutes back), so the quickstart dashboard opens onto a live-looking
+  session -- an amber stop-suspicion banner and populated liveness
+  rows -- instead of events aged out of the 24h scope. Judgements
+  remain time-independent: every offset clears the 15-minute stopped
+  threshold from the moment of generation.
+- Version single-sourcing: the CLI banner now reads
+  `agent_log_verifier.__version__` instead of a hardcoded "0.2", and
+  `__version__` (which had drifted at 0.1.0) now matches
+  `pyproject.toml`.
+- Public-release housekeeping: `scripts/agent-monitor.bat` picks the
+  most recently modified project directory under
+  `%USERPROFILE%\.claude\projects` (or takes one as an argument)
+  instead of hardcoding the author's machine path; GitHub Actions CI
+  mirrors the local gate (pytest / mypy --strict / ruff) on
+  Ubuntu+Windows x Python 3.12/3.13; a tag-triggered release workflow
+  publishes to PyPI via Trusted Publishing (no stored token).
+
+### Added
+- False-test-report detection (COMPLIANCE-004): flags an assistant
+  claim that tests passed when no test-run command appears in the
+  prior 8 tool calls. Vocabulary picked from a 277-transcript real-log
+  mining pass: 「完了しました」 (290 hits) is a generic done phrase,
+  not a test claim, and is excluded to avoid becoming the biggest
+  false-positive source; 「verified」/「検証済」 are excluded as
+  multi-sense. v1 covers "claimed but never ran"; the "ran but the
+  result was a failure" variant is deferred.
+- Data-loss operation detection (INCIDENT-001): surfaces executed
+  data-loss commands (rm -rf, git reset --hard, git clean -f, plain
+  git push --force, git checkout --, DROP TABLE/DATABASE, terraform
+  destroy, Remove-Item -Recurse -Force, rmdir /s). Real-log mining
+  drove two carve-outs: taskkill //F (30 hits, dev-server restart, not
+  data loss) and git push --force-with-lease (the safe leased form) do
+  NOT flag. Reports an executed fact, not a proven violation --
+  authorization is left to the operator. Real-log scan: 14 findings,
+  zero taskkill false positives.
+- Loop detection (LOOP-001..004): consecutive error streaks, fruitless
+  identical retries, no-progress identical repeats, and
+  correction-ignored retries. All rules are error/result-anchored so
+  progress-making repetition (TDD re-runs, polling) never flags.
+  Grounded in a 277-transcript real-log mining pass plus an external
+  survey of loop causes.
+- Live dashboard (`alv serve`): 10s auto-refresh with UI-state
+  preservation, pause button, health banner, running spinners, task
+  labels as row identity (agent id demoted to a sub-line), plain-language
+  status tooltips, evidence-labelled completion verdicts.
+- User-adjustable tables: drag-resizable columns (double-click to
+  reset, persisted in localStorage across refreshes), rows-per-page
+  select (20/50/100/all), one-line ellipsis clipping for stable row
+  heights.
+- Zero-token completion auto-correction: background-agent
+  task-notification matching (task-id / tool-use-id keys) and
+  final-response-shape inference. Real-log impact: stopped_suspected
+  244 -> 6, unknown 76 -> 22.
+- Failure-mode capability matrix (internal docs): 20 modes classified
+  as implemented / implementable / conditional / impossible, with the
+  evidence-first design principles.
+- Scope-drift candidate presentation (COMPLIANCE-006): compares the
+  order prompt's mentioned-path set against actually-changed files and
+  reports, once per unit, changes falling outside every mentioned
+  path's project -- with the structural-approximation limitation stated
+  in the finding text itself (true adjacency is semantic). Mining over
+  1,850 units decomposed a naive 150-file out-of-scope tail into four
+  baked-in carve-outs: reference-only mentions (~70; Japanese puts the
+  reference verb AFTER the path, so a bidirectional window with a tight
+  save-verb override replaces the preceding-window check), temp/
+  scratchpad (24), .claude harness bookkeeping (17), doc rides-along
+  (45). No mentioned paths -> no scope basis -> silent. Real-log scan:
+  10 findings (demo-build spillover, unrequested infra, plus the
+  known prose-directed-bootstrap family recorded for observation).
+- Undisclosed-TODO completion detection (COMPLIANCE-005): a fresh
+  TODO/FIXME/NotImplemented marker added to a CODE file, followed in the
+  same user-turn window by a completion claim that does not mention the
+  leftover. Ships ahead of local evidence (mining: 2,171 Edit/Write
+  calls held ZERO such co-occurrences) on external grounding
+  (failure-mode matrix #12, fail-plausible study), with two
+  locally-evidenced guards: docs carve-out (9/10 real fresh insertions
+  were prose task lists) and a strict marker-count-increase requirement
+  so moving existing code never flags. TODO alone never fires;
+  disclosed leftovers (「TODO を残しています」) never fire. Real-log
+  scan: exactly 1 finding -- this feature's own mining script, the
+  documented fixture-vocabulary channel proving the wire-up end-to-end.
+- Error-breakdown dashboard tab (内訳): classifies every failed
+  tool_result into 10 categories (pre-read edit, missing path,
+  infra-guard hook, stale edit context, user rejection, shell syntax,
+  auto-mode denial, invalid arguments, old_string mismatch, other) and
+  shows category x count x share plus a per-tool table. Analysis, not
+  detection: the category keys were seeded from a 281-transcript mining
+  pass (601 error results, 231 unique prefixes) and anything
+  unrecognized lands in その他 honestly (35.9% at ship time -- the
+  visible size of that bucket is the backlog signal for extending the
+  dictionary). run_scan() now returns this tally as a third element,
+  computed in the same walk with no re-parse.
+
+### Fixed
+- Table column rules: with drag-resizable widths there was no visual
+  cue for where a column ends (operator request, 2026-07-12) -- cells
+  now carry a right border, last column open.
+- Liveness table row order: non-running rows kept discovery order
+  (agent-id glob order, effectively random by date), so a 44-day-old
+  completed agent could sit directly under the running rows and read
+  like the dashboard itself was stale (operator report, 2026-07-12).
+  Now running rows pin first, everything else sorts newest-first by
+  elapsed minutes; rows with no recency signal go last.
+- Path extraction: forward-slash drive form (`F:/...`) no longer loses
+  its drive letter; URLs, mid-word fragments, backtick enumerations and
+  single-segment fragments are no longer extracted as paths.
+- Prohibition extraction: keywords inside 「」/（） quotes are treated
+  as quoted names, not directives; action nouns must be adjacent to the
+  keyword; document writes (.md/.txt) under a "no implementation" order
+  are recognized as the ordered deliverable. Suspected-deviation count
+  on real logs: 194 -> 75.
+- Dashboard rule_id filter was never wired; rows now carry data-rule-id
+  and the selection survives fragment refreshes.
+- `/favicon.ico` returns 204 instead of a console-visible 404.
+- UnicodeEncodeError on legacy-codepage consoles: Windows defaults
+  stdout to a strict legacy codepage (cp932 on Japanese Windows), and a
+  single unencodable character -- first real-world hit: the CLI
+  banner's own em dash under a piped stdout -- aborted the whole scan.
+  The CLI now reconfigures stdout/stderr to errors="replace" at entry
+  (findings text is arbitrary transcript content), and the banner uses
+  an ASCII "--". File output was already explicit UTF-8 and is
+  unchanged.
+
+## [0.2.0] - 2026-07-11
+
+### Added
+- COMPLIANCE-002 (prohibited tool usage) and COMPLIANCE-003 (artifact
+  path match) detectors on top of v0.1's COMPLIANCE-001.
+- Constraint auto-extraction from order prompts (prohibitions, artifact
+  paths, scope boundaries).
+- Main-session turn segmentation: compliance runs per user turn instead
+  of first-prompt-only.
+- Static dashboard (`alv dashboard`) and live server (`alv serve`).
+- v1.0 publication kit: synthetic demo-log generator, English README,
+  MIT license, pip build verification.
+
+### Fixed
+- COMPLIANCE-001 false positives reduced 607 -> 39 (93.6%) via
+  confirmation-keyword windows, reference-section stripping, and an
+  artifact-extension whitelist.
+
+## [0.1.0] - 2026-07-11
+
+### Added
+- Crash-tolerant Claude Code JSONL parser with secret-guarding parse
+  warnings.
+- COMPLIANCE-001: illusory-save detection (a "saved it" claim with no
+  matching Write/Edit call).
+- LIVENESS-001: 4-value subagent liveness inference
+  (running/completed/stopped_suspected/unknown) with parent-session
+  correlation.
+- CLI: `alv scan | watch | report` with text/JSON/markdown output.

agent_log_verifier-0.3.0/CLAUDE.md

@@ -0,0 +1,67 @@
+# agent-log-verifier — agent instructions
+
+Post-hoc, read-only, zero-token log-analysis dashboard for Claude Code
+transcript logs (JSONL). Python 3.12+, pydantic models, stdlib http
+server. This is a log-reading tool for a human maintainer to skim in
+place of re-reading the JSONL by hand -- it is not a security tool,
+not an anomaly monitor, and not a violation catcher. Names like
+`COMPLIANCE-001` and `INCIDENT-001` in this codebase are internal
+category labels for log-event lookups; the `detectors/` directory
+name is kept for import compatibility.
+
+## Process (binding)
+
+Read [docs/DEVELOPMENT.md](docs/DEVELOPMENT.md) before non-trivial
+work. Short version:
+
+- Work starts as a [docs/BACKLOG.md](docs/BACKLOG.md) entry, not from
+  chat context alone.
+- New lookup rules need BOTH external grounding (research) AND a
+  real-log mining pass proving the pattern occurs locally. Record
+  evidence-based rejections too.
+- TDD: red test first. Gate: `powershell -File scripts/gate.ps1`
+  (pytest + mypy --strict + ruff; pre-commit hook enforces it).
+- Lookup changes: run `alv scan` on the live project directory and
+  put before/after numbers in the commit message.
+- Update CHANGELOG.md `[Unreleased]` in the same commit.
+- Conventional Commits.
+
+## Layout
+
+- `src/agent_log_verifier/parser/` — crash-tolerant JSONL parsing
+- `src/agent_log_verifier/detectors/` — Lookup plugins (directory
+  name kept for import compatibility):
+  `compliance.py` (COMPLIANCE-001/002/003/004), `completeness.py`
+  (COMPLIANCE-005), `scope.py` (COMPLIANCE-006), `loops.py`
+  (LOOP-001..004), `incident.py` (INCIDENT-001), `liveness.py`
+  (LIVENESS-001; orchestrated directly by cli.py, not via the
+  registry — it needs parent-session context)
+- `src/agent_log_verifier/extraction/constraints.py` — order-prompt
+  constraint extraction (do-not / artifact paths / scope)
+- `src/agent_log_verifier/report/html_output.py` — static + live
+  dashboard (single-file HTML, all CSS/JS inline, no external
+  requests); `report/breakdown.py` — failed-tool_result
+  categorization behind the dashboard's 内訳 (breakdown) tab
+- `src/agent_log_verifier/server.py` — `alv serve` (localhost-only)
+
+## Design principles (do not violate)
+
+- Read-only over the observed system; never call an LLM.
+- Error/result-anchored lookup: only failure and non-progress are
+  worth noting; healthy repetition must be structurally unable to
+  flag.
+- Structural matching only; no natural-language similarity in
+  lookup logic.
+- Honest limits: log patterns the tool cannot spot are documented,
+  not approximated.
+
+## Local quirks (Windows, this machine)
+
+- Restart `alv serve` after edits (Python loads code at start):
+  `netstat -ano | grep ":8799.*LISTENING"` → `taskkill //F //PID
+  <pid>` → relaunch → sleep ~14s for first scan.
+- Dashboard screenshots: headless Playwright (`node` + any local
+  playwright-core install; a sibling project's node_modules works).
+  The Claude Browser pane screenshot times out on this page.
+- Bash heredocs mangle backslashes/UTF-8 here — write files with
+  the dedicated file tools, not `cat <<EOF`.

agent_log_verifier-0.3.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Skova Digital
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.