agentproof-cli 0.3.0__tar.gz

agentproof_cli-0.3.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 AgentProof contributors
+
+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.

agentproof_cli-0.3.0/PKG-INFO

@@ -0,0 +1,158 @@
+Metadata-Version: 2.4
+Name: agentproof-cli
+Version: 0.3.0
+Summary: Local evidence, checkpoints, and verification for AI coding tasks
+License: MIT
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Dynamic: license-file
+
+# AgentProof
+
+**Evidence for AI coding tasks. Local checkpoints, real command results, honest handoff reports.**
+
+[中文说明](README.zh-CN.md) · [Five-minute start](docs/quickstart.md) · [Example report](examples/report.md) · [Changelog](CHANGELOG.md) · [Security](SECURITY.md) · MIT license
+
+AI says “done.” What actually changed? Which checks ran, and do they still match the code?
+AgentProof answers those questions without another model or an API key.
+
+```text
+$ agentproof init --goal "Fix payment retries"
+Task initialized. Baseline: ... (42 files)
+$ agentproof verify --kind test -- python -m unittest discover -s tests
+PASS (exit 0). Log: .agentproof/checks/....log
+$ agentproof report
+/path/to/project/.agentproof/report.md
+```
+
+## Install
+
+Requires Python 3.10+, Git, and macOS, Linux, or Windows.
+This project is not published to PyPI yet.
+
+With [pipx installed and on PATH](https://pipx.pypa.io/stable/installation/), install directly from GitHub:
+
+```bash
+pipx install 'git+https://github.com/bzwei58/agentproof.git@main'
+```
+
+Then run `agentproof --version` from any project. If the command is missing, run
+`pipx ensurepath` and reopen your terminal. The package is named `agentproof-cli`, while
+the command is `agentproof`; remove it with `pipx uninstall agentproof-cli`.
+This installs the current main branch, not a versioned release.
+
+If you do not use pipx, create a virtual environment and install the same Git URL with its
+Python: `python3 -m venv /your/chosen/venv`, then
+`/your/chosen/venv/bin/python -m pip install 'git+https://github.com/bzwei58/agentproof.git@main'`.
+Use `/your/chosen/venv/bin/agentproof` or activate that environment.
+
+Start with the [five-minute walkthrough](docs/quickstart.md): one failing test, a fix,
+a stale check and a report, all in a disposable directory. For a real project, follow the workflow below.
+
+## Workflow
+
+Run from any directory inside the Git repository you want to inspect. Commands execute at
+that repository's root. One active task is supported per checkout; use a worktree for another task.
+
+1. **Start before AI edits.** `agentproof init --goal "Fix payment retries"` captures initial
+   local changes too, without creating commits or touching the Git index.
+2. Let your existing coding agent work. AgentProof does not launch or intercept it.
+3. **Save a milestone.** `agentproof checkpoint --label "first implementation"`.
+4. **Run real checks.** `agentproof verify --kind test -- python -m unittest discover -s tests`.
+   Run each test, lint or build command separately. Shell operators are not interpreted;
+   explicitly use `sh -c '...'` if you need shell syntax. Only run commands you trust.
+5. **Optionally import session signals.** `agentproof import-session /path/to/session.jsonl`.
+6. **Review AI claims.** `agentproof review --summary ai-summary.txt` (one claim per line), or repeat
+   `--claim "All tests passed"`. It reports `OBSERVED`, `UNVERIFIED` and `CONFLICT`; natural-language claims are not certified.
+7. **Review.** `agentproof status` prints evidence; `agentproof report` writes Markdown.
+8. **Recover selected snapshot contents.** `agentproof restore CHECKPOINT_ID` previews changes;
+   add `--apply` to write them, after an automatic backup checkpoint.
+
+`verify --timeout 60 -- COMMAND ...` sets a timeout in seconds (default 300).
+A nonzero command result returns CLI exit code 1; usage/operational errors return 2.
+Command stdout and stderr are retained in `.agentproof/checks/`. On failure, the terminal
+also shows the last 12 lines, capped at 4,000 characters, plus an explicit timeout message
+when applicable. Successful commands stay quiet. The full log is always retained locally.
+
+## What is recorded
+
+- SHA-256 fingerprints and modes of captured files, plus full file checkpoints.
+- Added, modified and deleted files relative to the task's initial checkpoint.
+- Actual verification command, exit status, output log and before/after fingerprints.
+- Whether each check still matches the current captured files. Editing code after a check
+  or having a check modify files marks it **STALE**.
+- Long or multiline commands have compact report summaries; expand **Full command** to
+  inspect the exact recorded invocation.
+- Optional Claude Code JSONL counts: tool calls, error results, compaction boundaries,
+  truncation markers and malformed records. Raw imported messages are not copied.
+- Claim review matches completion statements to local file changes and fresh checks using transparent rules.
+
+## Boundaries
+
+This is an **evidence recorder and claim-gap detector**, not an automatic correctness judge. A passing check does
+not prove requirements satisfaction, coverage or absence of bugs. There is no model-based
+risk scoring, live session watcher, automatic test discovery or cross-agent correlation yet.
+Claim review never calls a model and cannot decide whether a requirement is substantively correct.
+Session import is best-effort for Claude Code-style `message.content` blocks and
+`system/compact_boundary` records. Other formats and future schema changes may not be counted.
+Truncation text is a heuristic signal, not proof that important context was lost.
+
+Snapshots include Git-tracked and non-ignored untracked regular files. Ignored files,
+symlinks, submodule contents, Git index state, external services and environment variables
+are outside the captured scope. Known secret filenames (`.env*`, private-key extensions,
+`credentials.json`) are excluded and listed; **this is not secret detection**. Verification
+logs can contain secrets printed by your own commands. Keep `.agentproof/` private.
+Initialization adds a local Git exclusion, and evidence is stored with a private directory mode.
+
+A file over 10 MiB or total captured content over 100 MiB aborts snapshotting. Avoid concurrent
+edits while capturing or verifying. Only one AgentProof command can run per checkout.
+After an interrupted process, inspect and remove a stale `.agentproof/lock` manually.
+
+Restore rewrites files present in the target snapshot. **Files added later remain**; this is
+file recovery, not an exact whole-workspace rollback. It leaves Git history and staging alone.
+It rejects symlink paths, verifies snapshot content hashes and preflights file/directory
+conflicts, including blocked parent paths, before changing files.
+It is not a filesystem transaction; concurrent changes or I/O errors can still interrupt writes.
+Restore does not recover excluded files or delete newly created ones. A pre-restore checkpoint lets you
+recover overwritten captured content. Do not manually edit snapshot manifests or files.
+
+## Development
+
+```bash
+git clone https://github.com/bzwei58/agentproof.git
+cd agentproof
+python3 -m venv .venv
+.venv/bin/python -m pip install -e .
+python3 -m unittest discover -s tests -v
+python3 -m agentproof --help
+```
+
+Want to try it on two real tasks? See the [pilot guide and feedback form](docs/pilot.md).
+
+Tests use disposable Git repositories and cover stale evidence, nonzero exits, timeout,
+initial uncommitted edits, restoration backups, symlink rejection, secret exclusions and
+partial session logs. GitHub Actions runs them on macOS/Linux and Python 3.10/3.12/3.13.
+
+Ideas for later: task history, more session adapters, configurable snapshot scope, optional
+CI handoff artifacts. Contributions should preserve the distinction between observed facts
+and model-generated claims. Please include a reproducible example and tests for behavior changes.
+
+## Evidence matching in 0.3.0
+
+Declare each check with `verify --kind test|lint|build|other -- COMMAND`. This is a
+user-supplied category, not a certification. Old records remain usable as command
+history but do not acquire a check category automatically.
+
+Natural-language completion claims are never certified as `VERIFIED`. `OBSERVED`
+means a corresponding category of command exited 0 with an intact log and matching
+captured files; inspect the named command and output to confirm scope and test count.
+A build cannot support a test claim. Missing, stale, failed, or modified evidence leaves
+the claim `UNVERIFIED`. Reviews are recalculated on every `status` or `report` call.
+Already exported Markdown is a static snapshot and must be regenerated after changes.
+External services, ignored dependencies and environment changes remain outside the fingerprint.
+
+Session import recognizes Claude message blocks and Codex rollout `response_item.payload`
+tool events plus `compacted` records. Opaque outputs do not establish command success.
+Cline and Cursor native formats are not yet validated; generic top-level event counting
+must not be interpreted as full client support.

agentproof_cli-0.3.0/README.md

@@ -0,0 +1,148 @@
+# AgentProof
+
+**Evidence for AI coding tasks. Local checkpoints, real command results, honest handoff reports.**
+
+[中文说明](README.zh-CN.md) · [Five-minute start](docs/quickstart.md) · [Example report](examples/report.md) · [Changelog](CHANGELOG.md) · [Security](SECURITY.md) · MIT license
+
+AI says “done.” What actually changed? Which checks ran, and do they still match the code?
+AgentProof answers those questions without another model or an API key.
+
+```text
+$ agentproof init --goal "Fix payment retries"
+Task initialized. Baseline: ... (42 files)
+$ agentproof verify --kind test -- python -m unittest discover -s tests
+PASS (exit 0). Log: .agentproof/checks/....log
+$ agentproof report
+/path/to/project/.agentproof/report.md
+```
+
+## Install
+
+Requires Python 3.10+, Git, and macOS, Linux, or Windows.
+This project is not published to PyPI yet.
+
+With [pipx installed and on PATH](https://pipx.pypa.io/stable/installation/), install directly from GitHub:
+
+```bash
+pipx install 'git+https://github.com/bzwei58/agentproof.git@main'
+```
+
+Then run `agentproof --version` from any project. If the command is missing, run
+`pipx ensurepath` and reopen your terminal. The package is named `agentproof-cli`, while
+the command is `agentproof`; remove it with `pipx uninstall agentproof-cli`.
+This installs the current main branch, not a versioned release.
+
+If you do not use pipx, create a virtual environment and install the same Git URL with its
+Python: `python3 -m venv /your/chosen/venv`, then
+`/your/chosen/venv/bin/python -m pip install 'git+https://github.com/bzwei58/agentproof.git@main'`.
+Use `/your/chosen/venv/bin/agentproof` or activate that environment.
+
+Start with the [five-minute walkthrough](docs/quickstart.md): one failing test, a fix,
+a stale check and a report, all in a disposable directory. For a real project, follow the workflow below.
+
+## Workflow
+
+Run from any directory inside the Git repository you want to inspect. Commands execute at
+that repository's root. One active task is supported per checkout; use a worktree for another task.
+
+1. **Start before AI edits.** `agentproof init --goal "Fix payment retries"` captures initial
+   local changes too, without creating commits or touching the Git index.
+2. Let your existing coding agent work. AgentProof does not launch or intercept it.
+3. **Save a milestone.** `agentproof checkpoint --label "first implementation"`.
+4. **Run real checks.** `agentproof verify --kind test -- python -m unittest discover -s tests`.
+   Run each test, lint or build command separately. Shell operators are not interpreted;
+   explicitly use `sh -c '...'` if you need shell syntax. Only run commands you trust.
+5. **Optionally import session signals.** `agentproof import-session /path/to/session.jsonl`.
+6. **Review AI claims.** `agentproof review --summary ai-summary.txt` (one claim per line), or repeat
+   `--claim "All tests passed"`. It reports `OBSERVED`, `UNVERIFIED` and `CONFLICT`; natural-language claims are not certified.
+7. **Review.** `agentproof status` prints evidence; `agentproof report` writes Markdown.
+8. **Recover selected snapshot contents.** `agentproof restore CHECKPOINT_ID` previews changes;
+   add `--apply` to write them, after an automatic backup checkpoint.
+
+`verify --timeout 60 -- COMMAND ...` sets a timeout in seconds (default 300).
+A nonzero command result returns CLI exit code 1; usage/operational errors return 2.
+Command stdout and stderr are retained in `.agentproof/checks/`. On failure, the terminal
+also shows the last 12 lines, capped at 4,000 characters, plus an explicit timeout message
+when applicable. Successful commands stay quiet. The full log is always retained locally.
+
+## What is recorded
+
+- SHA-256 fingerprints and modes of captured files, plus full file checkpoints.
+- Added, modified and deleted files relative to the task's initial checkpoint.
+- Actual verification command, exit status, output log and before/after fingerprints.
+- Whether each check still matches the current captured files. Editing code after a check
+  or having a check modify files marks it **STALE**.
+- Long or multiline commands have compact report summaries; expand **Full command** to
+  inspect the exact recorded invocation.
+- Optional Claude Code JSONL counts: tool calls, error results, compaction boundaries,
+  truncation markers and malformed records. Raw imported messages are not copied.
+- Claim review matches completion statements to local file changes and fresh checks using transparent rules.
+
+## Boundaries
+
+This is an **evidence recorder and claim-gap detector**, not an automatic correctness judge. A passing check does
+not prove requirements satisfaction, coverage or absence of bugs. There is no model-based
+risk scoring, live session watcher, automatic test discovery or cross-agent correlation yet.
+Claim review never calls a model and cannot decide whether a requirement is substantively correct.
+Session import is best-effort for Claude Code-style `message.content` blocks and
+`system/compact_boundary` records. Other formats and future schema changes may not be counted.
+Truncation text is a heuristic signal, not proof that important context was lost.
+
+Snapshots include Git-tracked and non-ignored untracked regular files. Ignored files,
+symlinks, submodule contents, Git index state, external services and environment variables
+are outside the captured scope. Known secret filenames (`.env*`, private-key extensions,
+`credentials.json`) are excluded and listed; **this is not secret detection**. Verification
+logs can contain secrets printed by your own commands. Keep `.agentproof/` private.
+Initialization adds a local Git exclusion, and evidence is stored with a private directory mode.
+
+A file over 10 MiB or total captured content over 100 MiB aborts snapshotting. Avoid concurrent
+edits while capturing or verifying. Only one AgentProof command can run per checkout.
+After an interrupted process, inspect and remove a stale `.agentproof/lock` manually.
+
+Restore rewrites files present in the target snapshot. **Files added later remain**; this is
+file recovery, not an exact whole-workspace rollback. It leaves Git history and staging alone.
+It rejects symlink paths, verifies snapshot content hashes and preflights file/directory
+conflicts, including blocked parent paths, before changing files.
+It is not a filesystem transaction; concurrent changes or I/O errors can still interrupt writes.
+Restore does not recover excluded files or delete newly created ones. A pre-restore checkpoint lets you
+recover overwritten captured content. Do not manually edit snapshot manifests or files.
+
+## Development
+
+```bash
+git clone https://github.com/bzwei58/agentproof.git
+cd agentproof
+python3 -m venv .venv
+.venv/bin/python -m pip install -e .
+python3 -m unittest discover -s tests -v
+python3 -m agentproof --help
+```
+
+Want to try it on two real tasks? See the [pilot guide and feedback form](docs/pilot.md).
+
+Tests use disposable Git repositories and cover stale evidence, nonzero exits, timeout,
+initial uncommitted edits, restoration backups, symlink rejection, secret exclusions and
+partial session logs. GitHub Actions runs them on macOS/Linux and Python 3.10/3.12/3.13.
+
+Ideas for later: task history, more session adapters, configurable snapshot scope, optional
+CI handoff artifacts. Contributions should preserve the distinction between observed facts
+and model-generated claims. Please include a reproducible example and tests for behavior changes.
+
+## Evidence matching in 0.3.0
+
+Declare each check with `verify --kind test|lint|build|other -- COMMAND`. This is a
+user-supplied category, not a certification. Old records remain usable as command
+history but do not acquire a check category automatically.
+
+Natural-language completion claims are never certified as `VERIFIED`. `OBSERVED`
+means a corresponding category of command exited 0 with an intact log and matching
+captured files; inspect the named command and output to confirm scope and test count.
+A build cannot support a test claim. Missing, stale, failed, or modified evidence leaves
+the claim `UNVERIFIED`. Reviews are recalculated on every `status` or `report` call.
+Already exported Markdown is a static snapshot and must be regenerated after changes.
+External services, ignored dependencies and environment changes remain outside the fingerprint.
+
+Session import recognizes Claude message blocks and Codex rollout `response_item.payload`
+tool events plus `compacted` records. Opaque outputs do not establish command success.
+Cline and Cursor native formats are not yet validated; generic top-level event counting
+must not be interpreted as full client support.

agentproof_cli-0.3.0/agentproof/__init__.py

@@ -0,0 +1,2 @@
+"""AgentProof: evidence for AI coding tasks."""
+__version__ = "0.3.0"

agentproof_cli-0.3.0/agentproof/__main__.py

@@ -0,0 +1,3 @@
+from .cli import main
+
+raise SystemExit(main())