English | Chinese

PawnLogic

PawnLogic is a terminal-first autonomous AI agent with multi-provider model routing, persistent memory, real local tool execution, MCP integration, and a CTF-oriented toolchain. The current public release is 0.3.13.

System Requirements

• Linux or WSL2
• Python 3.10+
• pip
• git only for source checkouts, development, or git-backed skill packs
• ~/.local/bin in PATH when using the global pawn launcher
• Optional: Docker for container tools, browser dependencies for Patchright / Scrapling, and CTF packages for pwn workflows

Quick Start

Option A: install from PyPI

pip install pawnlogic
pawn

The first run opens the API key configuration flow. Runtime files are created under ~/.pawnlogic/, not inside the project directory.

Option B: one-line installer

curl -fsSL https://raw.githubusercontent.com/john0123412/PawnLogic/main/install.sh | bash
pawn

The installer creates an isolated venv under ~/.local/share/pawnlogic, installs the official PyPI package, and writes ~/.local/bin/pawn.

Option C: source checkout for development

git clone https://github.com/john0123412/PawnLogic.git
cd PawnLogic
python3 -m venv venv
source venv/bin/activate
pip install -e ".[dev]"
pawn

Optional extras:

pip install "pawnlogic[docker]"    # Docker SDK integration
pip install "pawnlogic[browser]"   # Scrapling + Patchright browser tools
pip install "pawnlogic[ctf]"       # pwntools, ROPgadget, ropper
pip install -e ".[dev,ctf]"        # source checkout with tests and CTF tools

pawnlogic[ctf] installs CTF tooling dependencies only. CTF skill packs are optional extension assets that users install explicitly, for example with /skills install <repo_url> into ~/.pawnlogic/skills. Third-party skill packs are not bundled into PyPI distributions unless their upstream license and notices have been reviewed for redistribution. Skill-pack manifests are runtime discovery metadata only; they do not authorize redistribution without a matching THIRD_PARTY_NOTICES.md entry. Git-backed skill-pack installs accept only https://, ssh://, or git@host:owner/repo.git remotes.

Source-checkout launcher fallback:

./pawn.sh

CLI entry points:

pawn
pawn --debug
pawn --eval "summarize this repository"
pawn --eval "summarize this repository" --json
pawn --continue                    # load the newest recoverable draft
pawn resume <session>              # load a chosen session without running it
python -m pawnlogic --help

Default pawn uses user-friendly output and hides raw tool-call internals, parser diagnostics, detailed reasoning streams, and low-level API errors. Tool-call recovery is best effort, not a guarantee. If a malformed tool-call attempt cannot be parsed, it produces no tool call and is not executed; detailed parser diagnostics remain hidden in user-friendly mode. Use pawn --debug or /mode when you need detailed diagnostics. With --json, each line is an independent NDJSON record. Existing text, chunk, and json records remain stable; versioned Agent lifecycle records use the additive {"type":"event","data":{...}} envelope.

What's New

Version 0.3.12 fixes a high-risk confirmation prompt that could freeze the session:

• An expired confirmation no longer leaves the terminal stuck: a high-risk approval prompt whose wait ran out used to stay mounted for the rest of the session. The live selector kept the keyboard, the composer went read-only, and every keystroke was swallowed as selector input — the reported "tool stage" stall, which only a force-quit cleared. The event loop that mounted the prompt now owns its deadline, and the tool watchdog reclaims a prompt its abandoned worker was blocked on.
• High-risk operations deny by default: the approval prompt opens pre-selected on Deny. Press y to approve, n / Esc / Ctrl+C to reject; a bare Enter rejects. Previously the prompt opened on "Approve and run" and the live terminal routed Enter to it, so an Enter meant for the composer could silently approve a high-risk command. The prompt also ignores every key until it has been painted once, so an unseen prompt can never resolve a keystroke.
• A waiting tool is now visible: while an approval prompt is mounted, the status line reads ⚠ awaiting confirmation — Esc to review instead of the ordinary in-flight indicator, so a blocked tool is distinguishable from a working one.
• Bounded, configurable confirmation wait: the wait is confirmation_wait_sec (default 300 seconds), clamped below the active tool_watchdog_sec so the prompt always tears down while the tool waiting on it is still alive. Both deadlines used to be the literal 600.
• Scripted owner acceptance: tools/owner_acceptance_probe.py automates the checks that can be scripted (entry point starts; the terminal guard restores raw mode, alternate screen, cursor, mouse capture, and bracketed paste) and reports the four that genuinely need human eyes as manual rather than passing.

See CHANGELOG.md for the full release history.

Key Capabilities

Supported Models

PawnLogic ships with preconfigured model aliases. Only active providers with a configured API key are shown in /model and Tab completion.

Custom provider model descriptions come from ~/.pawnlogic/custom_providers.json. Re-running /provider update <name> refreshes selected models and writes English fallback descriptions for fetched models when the provider does not supply a useful description.

Delegated tasks automatically prefer an eligible fast worker when no model request is supplied; they do not automatically reuse the current conversation model. /worker lists every model currently visible through /model, including eligible custom-provider aliases. /agent policy can allow or deny aliases, select the default routing mode, and cap cost or concurrency. Explicit model requests are preferences: provider visibility, user policy, capability, and budget checks remain authoritative. Structured tasks and results carry task/parent IDs, deadlines, usage, and failure records. Shared orchestration budgets are reserved atomically, and cancellation is cooperative. The core orchestrator admits at most two workers; each concurrent child has a copied RuntimeContext, an isolated workspace, a bounded output collector, and a task-local cancellation token. Concurrent children may use only task-isolated file Tools. delegate_task remains a single-task compatibility Adapter: a policy value of max-concurrency=2 takes effect only for a supported batch caller and never causes implicit fan-out.

Provider Management

/provider                         # open the provider TUI
/provider add <name> <base_url> <ENV_KEY> [anthropic]
/provider fetch <name>            # fetch available models and select aliases
/provider update <name>           # re-fetch provider models
/provider activate <name>         # show selected provider models
/provider deactivate <name>       # hide provider models
/provider list                    # show provider and key status
/provider test <model>            # test connectivity for a model alias
/setkey                           # run key setup again
/keys                             # show configured key status

API keys are stored in ~/.pawnlogic/.env. Provider configs, model aliases, and descriptions are stored in ~/.pawnlogic/custom_providers.json without secret values. Provider setup does not write keys into shell startup files.

The interactive TUI also edits a provider in place. Open a provider's detail view and choose Edit Provider to correct its Base URL and Format; the save keeps the provider name, its API key, and its loaded models. Renaming is not offered there, because a rename must re-point every model entry and the key's environment variable and cannot be written atomically. Replace the key with Update API Key, which asks for the full value again and never displays the stored one.

Confirmation dialogs mark the focused button in text as well as colour, and ← → ↑ ↓ and Tab all move between them. Delete Provider opens a dialog that starts on Cancel, so Enter never deletes by accident.

The model list behind Fetch and Sync opens with an empty search box every time, so a query typed into an earlier list is never re-applied to the next one. Move with ↑ ↓ PageUp and PageDown; Space or Enter ticks the model under the cursor, a selects all, and c clears the selection. Press s to load the ticked models and stay in the list, or S to load them and close it — neither requires moving to the buttons first. The list is paged, and its three actions — Load Selected, Load & Close, and Cancel — sit after the last model, so press L to jump straight to them instead of walking down once per model. They mark the focused one in text rather than colour alone.

Fetch and Sync never send a chat request, so listing models costs nothing. They read the provider's free /v1/models listing and hide entries that report a non-text output modality; a provider that does not report capability metadata keeps all of its entries. /provider test <model> is also free: it checks the same listing, so it answers "is the base URL reachable and does this key work" without ever inferring. The trade-off of having no billable probe anywhere in the provider flow is that a model your key cannot actually use is no longer filtered out in advance — it fails as a normal API error the first time you use it.

Plain http:// provider endpoints are allowed for local relays and lab setups, but user-friendly mode prints a trust-boundary warning because requests and API keys are not protected by TLS.

Unstable custom providers can be tuned through environment variables in ~/.pawnlogic/.env: PAWNLOGIC_API_RETRY_MAX controls total request attempts including the first attempt, PAWNLOGIC_API_RETRY_AFTER_MAX caps provider Retry-After delays, and PAWNLOGIC_API_CONNECT_TIMEOUT, PAWNLOGIC_API_READ_TIMEOUT, and PAWNLOGIC_API_NONSTREAM_TIMEOUT tune connection and response wait times.

Quick Command Reference

/model <alias>                    # switch model
/mode                             # toggle user-friendly/debug output
/chat find <keyword>              # search all sessions
/think <prompt>                   # run one deeper reasoning turn
/compact                          # summarize and compact context
/undo [n]                         # roll back recent turns
/queue                            # advanced queue inspection; does not interrupt the active Turn
/queue clear                      # clear queued/recovered messages without interrupting a Turn
/queue resume                     # resume recoverable queued work
/queue remove <id>                # remove one queued message by stable ID
/queue steer <id>                 # convert a follow-up into a steer
/queue follow-up <id>             # convert a steer into a follow-up
/queue recall <id>                # prefill the editor without removing the message
/abort                            # interrupt the active Turn and clear queued/recovered work
/deep                             # full-power mode
/max                              # maximum mode with up to 100 tool-call iterations
/ultra                            # MAX limits with up to 150 tool-call iterations
/init_project [desc]              # initialize project state
/pwnenv                           # check CTF toolchain integrity
/ctf init <name>                  # start CTF workspace metadata
/ctf solved [flag]                # mark a confirmed CTF flag as solved
/ctf writeup                      # export a CTF writeup draft
/skills install <repo_url>         # install a git-backed skill pack
/skills                            # interactive TUI: toggle, sync, rescan
/extension list                   # list installed Extensions
/extension enable <name>          # explicitly enable an Extension
/extension disable <name>         # disable an Extension
/worker [alias|auto]              # inspect or set the preferred worker
/planguard [strict|advisory|status]  # no argument opens the mode selector; tiers default to advisory, strict is opt-in
/agent policy show                # inspect delegated-agent policy
/agent run <role> <objective>     # print a safe delegate_task request template

Run /help inside PawnLogic for the full command list.

Trust Boundary

PawnLogic is an agent execution tool, not a security sandbox. It intentionally executes real tools with the current user's permissions when you ask it to do so. Pattern filters, Docker boundaries, and capability profiles reduce accidents; they do not contain a determined attacker.

Web fetches and browser navigation evaluate HTTP(S) targets through the shared Network Policy before use. URLs are normalized; embedded credentials, cloud-metadata/internal targets, the reserved localhost namespace (including subdomains), and loopback, link-local, multicast, unspecified, or reserved addresses are denied. Private-network targets require explicit authorization, and non-interactive requests fail closed when confirmation would otherwise be required. Redirect destinations are normalized, resolved, and evaluated again before they are followed, including any target-scoped authorization. Model-generated Tool arguments cannot grant private-network authorization, and confirmed private targets bypass remote reader services.

Docker bridge/host networking and legacy uvx mcp-server-fetch startup use capability-only authorization because no concrete URL is available at the gate. Authorize Docker networking with allow_network=true or PAWNLOGIC_DOCKER_ALLOW_NETWORK=true; authorize the legacy MCP network install with allow_network_install=true or PAWNLOGIC_MCP_ALLOW_NETWORK_INSTALL=true. These approvals grant only the named capability; they are not URL-target approvals.

User-friendly mode prints explicit trust-boundary notices for host shell execution, Docker container exec, browser/network-capable tools, private network URL access, delegated sub-agents, and plaintext HTTP providers. Use pawn --debug when you need lower-level tool arguments and diagnostics. Docker file mounts are workspace-bound by default, including read-only mounts; outside read-only challenge files require explicit allow_host_read_mount.

Host shell execution now passes through an operation policy before subprocess startup. Low-risk commands run normally, medium-risk commands are classified for audit, high-risk commands require explicit interactive confirmation, and critical operations are denied by default. The confirmation modal opens pre-selected on Deny: press y to approve, n/Esc/Ctrl+C to reject, and a bare Enter rejects. Approval is never a side effect of a keystroke meant for the composer, and while the modal is mounted the status line shows ⚠ awaiting confirmation — Esc to review. Non-interactive execution, including pawn --eval, fails closed when a high-risk command would require confirmation. DANGEROUS_PATTERNS remains only one misuse/risk classifier; it is not a sandbox boundary and cannot stop a malicious local user.

Host shell execution is hard-bounded: on timeout the whole process group receives SIGTERM and then SIGKILL, and cleanup never waits forever even if a child becomes uninterruptible (for example a WSL2 kernel stall). Registered tools additionally run under a watchdog (tool_watchdog_sec, default 600 seconds): a tool call that exceeds the limit is abandoned with an ERROR result so the session continues instead of freezing. An abandoned background thread may keep running until the process exits. A high-risk confirmation waits confirmation_wait_sec (default 300 seconds, clamped to stay below tool_watchdog_sec); that deadline belongs to the terminal event loop that mounted the modal, and an abandoned tool thread's confirmation is reclaimed when the watchdog expires, so a timed-out prompt can never leave the modal mounted.

Optional Extensions

Python distributions may advertise Extension metadata through the pawnlogic.extensions entry-point group. PawnLogic can list installed Extensions without loading their code. Installation never enables an Extension automatically.

/extension list
/extension status [name]
/extension enable <name>
/extension disable <name>

Enabled names are stored under ~/.pawnlogic/extensions/enabled.json. Extension startup failures are isolated from core startup, and contribution name conflicts are rejected instead of overwriting built-in Tools or commands. Dependency-heavy or security-sensitive Extensions must remain independently packaged and published. The core wheel contains no pawnlogic_security package, security console script, or security dependency; installing such a distribution would still require explicit /extension enable <name> authorization.

MCP Tool Integration

For pip or one-line installer users, PawnLogic creates editable templates in ~/.pawnlogic/ on startup:

pawn
cp ~/.pawnlogic/mcp_configs.example.json ~/.pawnlogic/mcp_configs.json
# edit ~/.pawnlogic/mcp_configs.json and add keys with /setkey or ~/.pawnlogic/.env
pawn

For source checkout users, the repository template can also be copied directly:

cp mcp_configs.example.json ~/.pawnlogic/mcp_configs.json

Supported example MCP servers include Tavily search, Playwright browser automation, and a filesystem bridge. External fetch MCP is disabled in the example because uvx mcp-server-fetch may contact PyPI during startup; use PawnLogic's built-in fetch_url unless you explicitly enable that MCP server.

MCP subprocess stderr is written to ~/.pawnlogic/logs/mcp/<server>.stderr.log by default. Set top-level "debug_stderr": true in mcp_configs.json when you want raw MCP stderr on the console. PawnLogic advertises MCP roots for the current working directory and ~/.pawnlogic/workspace.

Data Layout

All runtime data and API keys are stored in ~/.pawnlogic/.

~/.pawnlogic/
├── .env                    # API keys
├── custom_providers.json   # user provider configs, no keys
├── mcp_configs.json        # MCP server declarations
├── pawn.db                 # sessions, messages, knowledge base
├── global_skills.md        # GSA skill archive
├── skills/                 # optional user-installed skill packs
├── sessions/               # per-session scratch directories (session_<id>/)
├── workspace/              # auto-named task directories plus by-name/ aliases
└── logs/                   # audit logs

The project directory contains no secrets and is safe to commit or share.

Examples

Add a third-party API

/provider add myrelay https://api.myrelay.com/v1/chat/completions MYRELAY_API_KEY
/provider fetch myrelay
/provider activate myrelay
/model <alias>

Vision analysis

Analyze screenshot ./screenshot.png, extract the code and fix the bug.

CTF Pwn

/model ds-v4-pro
Analyze ./challenge, use pwn_debug to inspect registers at main breakpoint.

FAQ

Q: /model doesn't show new models after adding a provider? A: Configure its key, run /provider fetch <name>, select models, then /provider activate <name>.

Q: Can I abbreviate a slash command? A: Yes. Type a unique prefix or subsequence such as /plg; Tab completion lists /planguard, and pressing Enter normalizes the command. Only a unique match is dispatched; ambiguous input lists its candidates and runs nothing. All registered built-in commands participate in both Prompt Toolkit and readline completion.

Q: How do I choose a plan-guard mode? A: Run /planguard (or /plg) in an interactive terminal, then use Up/Down or 1/2 and press Enter. Use /planguard advisory, /planguard strict, or /planguard status for explicit or non-interactive use. Advisory is the default; in strict mode the first two tool-call batches without a plan block still run with a correction, and the third such attempt is stopped before its tools execute.

Q: How does the live composer handle input while a Turn runs? A: In Prompt Toolkit mode, one persistent terminal keeps model and Tool output above a fixed bottom composer and status toolbar. Completed output lines are handed to the host terminal through Prompt Toolkit's safe terminal handoff, so native scrollback, mouse selection, and copy remain available while the Application is running. Consecutive submissions appear as muted queue rows immediately above the composer. Enter submits a steer for the next Tool safe point; if a text-only response finishes first, each unclaimed steer runs as a separate subsequent Turn. Alt+Enter queues one follow-up for natural completion. Esc interrupts the active Turn and immediately hands control to queued work when one exists; with no queued work, the interrupted prompt becomes an editable recovered draft. On an idle empty composer, Esc, Up, or Alt+Up joins queued/recovered work into an editable draft. /queue remains an advanced diagnostic and management command and never pauses the active Turn. Readline remains explicitly serial and buffers input until the Turn completes.

Q: What happens when I interrupt a running turn? A: Pawn waits for cooperative cancellation to settle. If queued work exists, Esc treats it as the new steer and continues with that direction without creating a duplicate recovered row. If the queue is empty, the interrupted prompt is prefilled as an editable recovered draft without rerunning it; press Enter to retry it once, or edit it then press Enter to replace it exactly once (including an edit beginning with /). /queue remove <id>, /queue clear, /queue steer <id>, /queue follow-up <id>, and /queue recall <id> remain available for advanced queue management. /abort interrupts the active Turn and clears all queued/recovered work; there is no separate --all form. After a restart, pawn --continue loads the newest interrupted, running, or failed session, and pawn resume <session> loads a chosen session. Both commands show the history and prefill the draft without running it automatically.

Q: Test Connection fails but fetch succeeds? A: They now read the same free /v1/models listing. Fetch walks every page and can fail on a page error or a bad listing body; Test Connection checks one request. A failure from one and not the other usually means a retryable error — run it again.

Q: Where are API keys stored? A: ~/.pawnlogic/.env — outside the project, never tracked by git.

Q: pawn says command not found? A: export PATH="$HOME/.local/bin:$PATH"

Q: Browser tools say a module is missing? A: pip install 'pawnlogic[browser]' then patchright install chromium.

Q: Does it support local Ollama models? A: Yes. /provider add, Base URL http://localhost:11434, leave key empty.

Documentation

Support

• GitHub: github.com/john0123412/PawnLogic
• Issues: use GitHub Issues for bugs and feature requests.