add-cli-provider

name	add-cli-provider
description	Step-by-step guide for adding a new AI CLI backend (provider) to Leap, such as a new coding assistant. Covers the CLIProvider Strategy pattern, state detection, input protocol, menu handling, configure_hooks and hooks_installed, and registry wiring, including custom CLI variants of the five base CLIs. Use when adding, implementing, or registering a new CLI provider.

Add a New CLI Provider

Guide for adding a new AI CLI backend to Leap (e.g., a new coding assistant).

Leap uses a Strategy pattern — each CLI backend implements the CLIProvider abstract class. The provider defines identity, state detection, input protocol, menu handling, and hook configuration. The rest of the system (server, client, monitor, state tracker) uses the provider interface generically.

Key Constants & Enums

These are the global constants you should use throughout the codebase — never hardcode provider names or state strings:

# Provider registry (cli_providers/registry.py)
from leap.cli_providers.registry import DEFAULT_PROVIDER, get_provider, list_providers
DEFAULT_PROVIDER  # = 'claude' — used as fallback when provider name is missing

# State enums (cli_providers/states.py) — extend str for JSON transparency
from leap.cli_providers.states import AutoSendMode, CLIState

CLIState.IDLE             # == 'idle'
CLIState.RUNNING          # == 'running'
CLIState.NEEDS_PERMISSION # == 'needs_permission'
CLIState.NEEDS_INPUT      # == 'needs_input'
CLIState.INTERRUPTED      # == 'interrupted'

AutoSendMode.PAUSE        # == 'pause'
AutoSendMode.ALWAYS       # == 'always'

# Pre-built frozen sets for common checks
from leap.cli_providers.states import WAITING_STATES, SIGNAL_STATES, PROMPT_STATES

Rules:

• Use CLIState.IDLE instead of 'idle' in comparisons and dict keys
• Use AutoSendMode.PAUSE instead of 'pause' in comparisons and defaults
• Use DEFAULT_PROVIDER instead of 'claude' for fallback defaults
• Use WAITING_STATES instead of ('needs_permission', 'needs_input', 'interrupted')
• The socket protocol uses cli_state (not claude_state) and cli_running (not claude_running)

Overview of Touchpoints

Adding a new CLI provider requires changes in these areas:

1. Provider class — The core implementation (cli_providers/)
2. Registry — Register the new provider (cli_providers/registry.py)
3. Package exports — Update __init__.py exports
4. Hook configuration — How the CLI reports state changes to Leap
5. Shell launcher — Optional per-CLI shortcut script
6. Makefile — Hook cleanup on uninstall
7. ASCII banner — Automatically handled (uses display_name)
8. Monitor table — Automatically handled (uses display_name)
9. CLI selector — Automatically handled (reads from registry)
10. Shell flags — Automatically handled (generated from registry)
11. Documentation — Update CLAUDE.md and README.md

Step-by-Step

1. Create the Provider Class

Create src/leap/cli_providers/<name>.py inheriting from CLIProvider.

"""
<Display Name> CLI provider.

Implements the CLIProvider interface for <CLI tool description>.
"""

import json
import re
import time
from pathlib import Path
from typing import Any, Optional

from leap.cli_providers.base import CLIProvider


class <Name>Provider(CLIProvider):
    """Provider for <CLI tool> (<TUI type>, <language>)."""

    # -- Identity --------------------------------------------------------

    @property
    def name(self) -> str:
        return '<name>'  # lowercase, used in config/metadata

    @property
    def command(self) -> str:
        return '<binary>'  # binary name in PATH (e.g. 'mycli')

    @property
    def display_name(self) -> str:
        return '<Display Name>'  # human-readable (e.g. 'My CLI Tool')

    # -- State detection patterns ----------------------------------------

    @property
    def interrupted_pattern(self) -> bytes:
        # Byte string that appears in ANSI-stripped PTY output when interrupted.
        # Run the CLI, press Ctrl+C/Escape, and observe what text appears.
        return b'<pattern>'

    @property
    def dialog_patterns(self) -> list[bytes]:
        # Compact patterns (ANSI-stripped, spaces removed) that indicate
        # a permission/question dialog. ALL must be present for a match.
        # Return [] to disable PTY-based dialog detection (rely on hooks).
        #
        # To find these: run the CLI, trigger a permission dialog, then
        # examine the PTY output after stripping ANSI codes and spaces.
        return [b'<pattern1>', b'<pattern2>']

    # -- Hook configuration ----------------------------------------------

    @property
    def hook_config_dir(self) -> Path:
        # Directory where the hook script will be installed.
        return Path.home() / '.<cli_config_dir>'

    @property
    def requires_binary_for_hooks(self) -> bool:
        # Return True if this CLI is optional (hooks skipped if not installed).
        # Return False if this CLI should always have hooks configured.
        return True

    def configure_hooks(self, hook_script_path: str) -> None:
        """Install hooks into the CLI's configuration file."""
        # See ClaudeProvider or CodexProvider for reference implementations.
        # Key responsibilities:
        # 1. Load the CLI's config file (JSON, TOML, YAML, etc.)
        # 2. Remove any old Leap hook entries (marker: "leap-hook.sh")
        # 3. Add new entries that call hook_script_path with state args
        # 4. Write the config back ATOMICALLY (use leap.utils.atomic_write)

    def hooks_installed(self) -> bool:
        """True iff Leap's hooks are wired up for this CLI."""
        # Mirror image of configure_hooks(). Both halves must be true:
        # 1. self.hook_config_dir / "leap-hook.sh" exists on disk
        # 2. The CLI's settings file references "leap-hook.sh" from any
        #    hook entry. Wrap parse in try/except — corrupt or missing
        #    files return False (do NOT raise).
        # See ClaudeProvider or CodexProvider for reference impls.
        ...

Required Properties (Abstract)

These MUST be implemented — the class won't instantiate without them:

Property	Type	Purpose
name	str	Short ID for config/metadata (e.g. 'claude', 'codex')
command	str	Binary name to find in PATH
display_name	str	Human-readable name for UI
interrupted_pattern	bytes	Text indicating user interrupted the CLI
dialog_patterns	list[bytes]	Patterns indicating a permission/input dialog
hook_config_dir	Path	Directory for hook script installation
configure_hooks()	method	Installs hooks into CLI config (use atomic writes)
hooks_installed()	method	Returns True iff Leap's hooks are currently wired up — used by the session-start gate to refuse to spawn the server when integration is missing (e.g. CLI installed after Leap). Mirror image of configure_hooks()

Optional Properties (Have Defaults)

Override these only if the CLI differs from the defaults:

Property	Default	When to Override
trust_dialog_patterns	Claude's trust dialog	Different startup dialog, or [] if no trust dialog
output_triggers_running	True	Set False for full-screen TUIs (Ratatui) where redraws look like output
enter_triggers_running	False	Set True for full-screen TUIs where Enter is the submit signal
silence_timeout	None (uses 15s global)	Shorter timeout for TUIs that output constantly during processing
has_numbered_menus	True	Set False if the CLI uses y/n prompts instead
menu_option_regex	None	Regex with groups (number, label) for numbered menus
free_text_option_prefix	None	Label prefix for "type your answer" options
below_separator_option_prefix	None	Label prefix for options needing arrow-key nav
paste_settle_time	0.15	Adjust if the CLI needs more/less time after paste
single_settle_time	0.05	Adjust for single-line input settle
image_prefix	'@'	Change if CLI uses different image attachment syntax
supports_image_attachments	False	Set True if CLI supports inline image files
requires_binary_for_hooks	False	Set True if hooks should only configure when CLI is installed
base_type	self.name	For built-in providers, leave the default — it returns the provider's own name. Custom providers (CustomCLIProvider) inherit the value from their wrapped base automatically via __getattribute__ delegation; you don't write base_type yourself. The session-start gate uses get_provider(provider.base_type).hooks_installed() so custom CLIs share their base's hook setup. All custom CLIs must wrap one of the five base CLIs — there is no path for a custom CLI that's not a variant of a built-in.
valid_signal_states	SIGNAL_STATES	Override if the CLI writes different states to signal files
supports_resume	False	Set True when you wire up the Leap Resume feature (see below)
requires_cwd_bound_resume	False	Set True if resuming this CLI requires running from the recorded cwd (see Cross-cwd resume — the "move" mechanism below). Drives the picker's Original / Current prompt.
supports_context_usage	False	Set True when you implement context_usage(). False renders the monitor's Context cell as N/A (CLI can't report usage, e.g. Cursor); True shows the % or blank. See Context-usage column below.

Optional Methods (Have Defaults)

Method	Default Behavior	When to Override
send_message()	Write text + settle + CR	Custom input protocol (e.g. char-by-char for raw mode)
send_image_message()	Same as send_message()	CLI has special image confirmation flow
is_image_message()	Check supports_image_attachments + prefix	Different image detection logic
select_option()	Returns error	Implement for numbered menus, y/n prompts, etc.
send_custom_answer()	Returns error	Implement for free-text input in dialogs
find_cli()	Searches PATH for self.command	Custom binary location logic
get_spawn_env()	Sets LEAP_TAG, LEAP_SIGNAL_DIR, LEAP_PYTHON, LEAP_CLI_PROVIDER	Additional env vars needed by the CLI
parse_signal_file()	Parses JSON {"state": "..."}	Different signal file format
extract_session_id()	Returns None (no resume)	Implement for Leap Resume — pull the session id out of the hook payload
resume_args()	Returns []	Implement for Leap Resume — return the argv tokens that resume the given session id
relocate_session()	Returns None (no cross-cwd)	Implement for the move mechanism — physically (or logically) bring the session's on-disk state under the user's chosen cwd. Required when requires_cwd_bound_resume = True.
session_exists()	Returns True	Override if your CLI records sessions with empty transcript_path so the picker's path-based stale-check can't filter them — return False when the session's on-disk state has been deleted out-of-band.
context_usage(cli_name, tag, storage_dir)	Returns None	Implement (with supports_context_usage = True) to show the monitor's Context column — locate your own source (a transcript via latest_transcript_for, or Copilot's status-line state file) and return prompt-tokens vs the context window. See Context-usage column below.
session_cost(cli_name, tag, storage_dir)	Returns None	Implement (with supports_cost = True) to add cost lines to the Context tooltip — sum the session's billable tokens from your transcript and price them via leap.utils.pricing (price_for / cost_usd). See Context-usage column + cost_usage.py.
input_history(cwd)	Returns None (passthrough — CLI handles ↑/↓ natively)	Return the CLI's on-disk input history (oldest→newest) so Leap drives ↑/↓ recall and keeps its input mirror in sync for ^^. Opting in steals ↑/↓ — read the ↑/↓ recall, dialogs & 'Last Msg' note below.
extract_last_user_prompt(cwd, tag, storage_dir, cli_name='')	Returns '' (monitor falls back to Leap's PTY recently_sent capture)	Read the user's most recent prompt from the transcript so the Last Msg column is accurate. The PTY fallback can carry stray echoed keystrokes (e.g. a leading 2 → "2hi"); reading the transcript avoids that. Resolve via latest_transcript_for(storage_dir, cli_name or self.name, tag) (cli_name is the recorded cli_sessions subdir, so a custom CLI built atop your provider resolves its own transcripts).
screen_shows_selection_dialog(display_lines)	Generic detector: a numbered ›/❯/▶ cursor, or a footer line carrying confirm/cancel/navigate hints (≥2 hints, or a · separator, or a short hint-only line) — cursor-glyph independent	Override only if your CLI's arrow-navigable dialog isn't caught by the generic detector. Drives ↑/↓ passthrough — see the note below.
deconfigure_hooks()	Removes leap-hook.sh + leap-hook-process.py from hook_config_dir	Override if configure_hooks() also writes into a CLI settings/config file — undo those changes surgically, write back atomically, then call super().deconfigure_hooks(). See deconfigure_hooks() section in step 4.

Leap Resume feature (leap --resume)

If this CLI supports resuming a previous conversation, implement the three resume hooks so the tag shows up in the leap --resume picker (prefixed with a [<display_name>] badge). All three must be set together:

1. supports_resume → True

2. extract_session_id(hook_data: dict) -> Optional[str] Given the JSON the CLI sends to leap-hook.sh on Stop / Notification events, return the stable session identifier (UUID / chat id / whatever your CLI uses). Return None when the payload isn't one of this CLI's sessions — the session recorder will then skip it.

   Examples: Claude derives it from transcript_path basename; Codex reads the session_id field directly (and falls back to the first JSONL line's payload.id).

3. resume_args(session_id: str) -> list[str] Return the argv tokens that, when prepended to the CLI invocation, resume the session. The server prepends these so positional subcommand forms stay in the right spot. Examples:

   # Claude: flag-value form, `=` is required so the single token
   # survives leap-server.py's argv pipeline intact
   return [f'--resume={session_id}']
   
   # Codex: positional subcommand
   return ['resume', session_id]
   
   # Cursor Agent (hypothetical): bare flag-value
   return ['--resume', session_id]
   

Data flow — no extra code is needed beyond these three methods:

• The hook (leap-hook-process.py) reads LEAP_CLI_PROVIDER (set by get_spawn_env) and calls your provider's extract_session_id with the raw hook payload. Matching sessions land in .storage/cli_sessions/<name>/<tag>.json.
• The picker scans .storage/cli_sessions/*/ and shows each tag as [<display_name>] <tag>. Custom CLIs appear automatically as long as they're registered.
• On selection, leap-resume.py sets LEAP_RESUME_SESSION_ID, LEAP_RESUME_CLI and LEAP_CLI, execs leap-main.sh, and leap-server.py consults your provider's resume_args before the PTY spawn.

If the session is tied to a specific working directory (Claude stores transcripts under a cwd-derived slug), record cwd in the hook payload — the picker chdirs there before launch so resume can find the transcript.

Gotchas observed in the wild:

• Some CLIs (Codex 0.121+) require a non-obvious schema — e.g. events nested under a top-level "hooks" key in hooks.json. If implementing configure_hooks, verify the resulting JSON actually triggers the hook by checking for an entry in .storage/logs/hook-debug.log (create the logs/ dir to enable).
• Some CLIs (Cursor Agent) gate hooks behind a server-side feature flag — on plans where the flag isn't enabled, the hook silently never fires regardless of schema validity. That's outside our control; implement the protocol anyway so users with the flag get the feature.
• Some CLIs (Codex) strip env vars when spawning hook subprocesses. leap-hook.sh already walks the PPID chain looking for a <project>/.storage/pid_maps/<pid>.json mapping — that mapping is written with cli_provider so the fallback still identifies the CLI. The project path itself is recovered from $LEAP_PROJECT_DIR or, if that's also been stripped, by regex-reading the install-time export LEAP_PROJECT_DIR="…" line out of ~/.zshrc / ~/.bashrc. You get this for free by using get_spawn_env (base class) without overriding it.

Cross-cwd resume — the "move" mechanism

When the user picks a session in leap --resume (or via the GUI's "From Resume" / "Open IDE + Move session" flows) from a different cwd than the one the session was originally recorded in, leap shows an arrow-key prompt:

  Where do you want to resume?
  ❯ CD into the original directory:  /Users/me/work/proj
    Stay in the current directory:   /Users/me

Both options must work for every CLI we ship. This requires every new resume-capable provider to implement the move mechanism:

1. requires_cwd_bound_resume → True This flips on the prompt above. When False, leap silently uses the current cwd (no prompt) — only correct for CLIs whose resume command finds sessions by id alone, regardless of cwd.

2. relocate_session(session_id, src_cwd, dst_cwd, *, transcript_path='', on_committed=None) -> Optional[str] Called when the user picks "Stay in the current directory". Must bring the session's on-disk state under dst_cwd so <cli> resume <id> finds it from there. Two flavors:

   • File-move (real) — for CLIs that store sessions in a cwd-derived path: physically move the transcript / chat dir across cwds. Use the shared primitives in src/leap/utils/relocation.py (signals_blocked, stage_copy_file/_tree, commit_file/_tree, verify_files_match, must_remove_tree, make_tmp_path). Wrap your orchestrator function in its own <name>_session_move.py next to the existing claude_session_move.py / gemini_session_move.py / cursor_session_move.py.

   • Logical no-op — for CLIs that key sessions by UUID alone (Codex): no files move. Just call on_committed(transcript_path) so leap's recorded cwd in .storage/cli_sessions/<name>/<tag>.json is bumped immediately, and return the unchanged transcript_path so the caller treats it as a successful relocation. Skip the file-move primitives and the signal-blocking — there's nothing critical to protect.

   Return value contract:

   • non-None string (the new path, or unchanged path for logical moves) → success, caller sets target_cwd = dst_cwd.
   • None → not applicable / can't be located; caller falls through to chdir into src_cwd (the "Original" path still works).
   • Raise RelocationError → real disk-side failure; caller surfaces the message to the user and exits non-zero. Source must be intact when this raises.

3. Reference behavior the four resume-capable built-in providers exhibit — pick the one your CLI most resembles and copy the shape:

   Provider	Storage layout	What relocate_session does
   ClaudeProvider	~/.claude/projects/<cwd-slug>/<uuid>.jsonl (+ optional <uuid>/ sidecar dir)	Atomic move of the JSONL and the sidecar tree across cwd-derived slugs. Pre-flight slug check, rogue-writer snapshot guards on both file and tree, rollback on sidecar-rename failure.
   GeminiProvider	~/.gemini/tmp/<slug>/chats/session-…jsonl + ~/.gemini/projects.json registry mapping cwd → slug	Locate src by parsing first-line sessionId (filename embeds only an 8-char prefix), claim a fresh dst slug via Gemini's exact slugify(basename(cwd)) algorithm with -N disambiguation, atomically update projects.json, roll back the file commit if the registry write fails.
   CursorAgentProvider	~/.cursor/chats/<MD5(workspace)>/<chatId>/ (whole directory tree)	Move the full chat dir across MD5 hash dirs. find_chat_dir first tries MD5(prefer_cwd) then falls back to scanning every project hash dir for the chatId — cursor's workspace-root walk may have hashed a parent of the recorded cwd. Snapshot-based rogue-writer guard + best-effort prune of the now-empty src project hash dir.
   CodexProvider	~/.codex/sessions/YYYY/MM/DD/rollout-<ts>-<uuid>.jsonl (date+UUID, cwd-agnostic)	No file move. Just calls on_committed(transcript_path) so leap's recorded cwd is bumped immediately. Also returns ['-C', os.getcwd(), 'resume', session_id] from resume_args so codex's own "Choose working directory to resume" prompt doesn't fire on top of leap's prompt.

4. session_exists(session_id, cwd) -> bool (only if your CLI's records have empty transcript_path) The picker's stale-record filter normally checks os.path.getsize(transcript_path) — if your CLI doesn't expose a transcript path (e.g. Cursor records transcript_path: ""), that check can't fire and stale records linger forever. Override session_exists to do a cheap on-disk check (e.g. Cursor's find_chat_dir scans ~/.cursor/chats/<hash>/<id>/); the picker will hide records that return False.

You get these for free — no code needed, but worth knowing:

• Records bookkeeping by session_id. relocate_records() in leap.utils.resume_store rewrites every cli_sessions/<cli>/<tag>.json entry matching a given session_id — not by transcript_path, which would silently no-op for empty-path records like Cursor's. The shared _on_committed callback in leap-resume.py calls it for you; you just need to invoke on_committed(new_path) from your relocate_session. Pass the new path for real moves, the unchanged path for logical no-op moves, or '' if your CLI doesn't track transcript paths.
• Hard-fail on dropped resume. When LEAP_RESUME_SESSION_ID is set but the resume can't be honored (unknown provider, no supports_resume, --cli mismatch, etc.), leap-server.py exits non-zero with a yellow ✗ Refusing to start stderr message instead of silently starting a fresh session — _apply_resume_or_fail handles this centrally. Just make sure your supports_resume accurately reflects whether relocate_session + resume_args are actually implemented.

TL;DR — minimum overrides for a new resume-capable CLI:

from typing import Any, Optional

class MyCLIProvider(CLIProvider):
    @property
    def supports_resume(self) -> bool:
        return True

    @property
    def requires_cwd_bound_resume(self) -> bool:
        # True for CLIs whose ``<cli> resume <id>`` only finds the
        # session when run from the recorded cwd; False for ones that
        # find sessions by id alone (e.g. Codex).
        return True

    def extract_session_id(self, hook_data: dict) -> Optional[str]:
        ...  # pull the session id from the hook payload

    def resume_args(self, session_id: str) -> list[str]:
        ...  # build the argv tokens that resume <session_id>

    def relocate_session(
        self,
        session_id: str,
        src_cwd: str,
        dst_cwd: str,
        *,
        transcript_path: str = '',
        on_committed: Optional[Any] = None,
    ) -> Optional[str]:
        # File-move flavor (like Claude/Gemini/Cursor):
        #   write src/leap/utils/<name>_session_move.py using the
        #   relocation.py primitives and call into it here.
        # Logical no-op flavor (like Codex):
        #   if on_committed is not None and transcript_path:
        #       on_committed(transcript_path)
        #   return transcript_path or None
        ...

    def session_exists(self, session_id: str, cwd: str) -> bool:
        # Only override if your CLI's records have empty
        # transcript_path (so the picker's path-based stale filter
        # can't see them).  Default returns True.
        ...

Context-usage column (the monitor's "Context" %)

The monitor's Context column shows, per session, how full the model's context window is (e.g. 43%) so the user sees how close it is to auto-compaction. It's opt-in per provider via two members:

• supports_context_usage (property, default False) — whether this CLI can report usage at all. False → the cell renders N/A (the CLI fundamentally can't, e.g. Cursor). True → the cell shows the % when available, or blank when context_usage() returns None (supported, but no data yet).
• context_usage(self, cli_name, tag, storage_dir) (default None) — the measurement. The provider locates its own source; cli_name is the row's recorded CLI name (pass it to latest_transcript_for so custom CLIs read their own cli_sessions/<name>/ subdir).

The monitor just calls get_provider(cli_provider).context_usage(cli_provider, tag, STORAGE_DIR) and renders ContextUsage.percent color-coded green→amber→red. You don't touch the monitor.

Two source shapes (add a public entry point in utils/context_usage.py):

1. Transcript CLIs (Claude / Codex / Gemini). Add a per-CLI _<cli>_usage_from_tail(tail: bytes) -> Optional[ContextUsage] and a thin <cli>_context_usage(path) = _context_usage(path, _<cli>_usage_from_tail) (shared stat/mtime cache + 32 KiB tail read + a final exception net — this runs on the render thread, so the parser must never throw; guard every field with isinstance). The provider resolves the path itself:

   @property
   def supports_context_usage(self) -> bool: return True
   def context_usage(self, cli_name, tag, storage_dir):
       tp = latest_transcript_for(storage_dir, cli_name, tag)
       return mycli_context_usage(tp) if tp else None
   

2. CLIs with no transcript usage but a status line (Copilot). Copilot's transcript exposes no live usage, but its status line receives the live numbers (current_context_tokens, context_window_size, model) on stdin each render. Leap installs leap-copilot-statusline.py (registered in ~/.copilot/settings.json by CopilotProvider.configure_hooks, chaining any existing status line via a leap-statusline-chain sidecar); the script writes <storage>/sockets/<tag>.context (the status-line subprocess inherits LEAP_TAG/LEAP_SIGNAL_DIR). The provider reads that file via copilot_context_usage(state_path). To add a status-line CLI: install its script in configure_hooks (and have _install_and_configure in configure_hooks.py copy it), keep hooks_installed() independent of it (the status line is optional — don't gate session startup on it), and read the state file in context_usage.

What to measure: the prompt size of the latest turn (the conversation loaded into the window), not the model's reply. Token semantics differ:

• Claude reports new (uncached) input separately from the cached prefix, so prompt = input_tokens + cache_creation_input_tokens + cache_read_input_tokens.
• Codex / Gemini / Copilot report input (or current_context_tokens) as the full prompt (cached is a subset), so just use it.

Window (the denominator): prefer a value the CLI records, and prefer the effective limit the CLI bases its own "% used" / compaction on so Leap's % matches what the user sees (Codex carries info.model_context_window; Copilot's status line carries displayed_context_limit - the limit it shows the user and auto-compacts against ~80% - preferred over the raw context_window_size). Otherwise map the model id to a window with a sane default (Gemini ~1M; Claude 200k, or 1M via the ~/.claude.json [1m] detection). Add a usage-overflow fallback if the window is ambiguous (usage above a window size proves a larger one).

When a CLI genuinely can't (leave supports_context_usage = False → N/A): Cursor — its CLI exposes no token usage at all, records no transcript_path, and stores chats in an opaque content-addressed SQLite blob store. There's no on-disk number to read, so its Context cell shows N/A.

↑/↓ history recall, dialogs & the "Last Msg" column

Three small features interact here; getting them right avoids a classic bug ("arrows stuck in a selection dialog").

1. input_history(cwd) — opt into ↑/↓ recall. Return the CLI's own on-disk history (oldest→newest) and Leap intercepts ↑/↓ to drive recall itself, keeping its input mirror in sync so a later ^^ captures the recalled text. Returning None = passthrough (the CLI handles ↑/↓).

2. The dialog trap. Once you implement input_history, Leap steals ↑/↓ at the prompt — including while your CLI is showing an arrow-navigable selection dialog (permission/trust/model picker). To stop that, the input filter passes ↑/↓ through whenever CLIStateTracker.screen_has_active_dialog() is True. That method now calls the generic CLIProvider.screen_shows_selection_dialog(display_lines) first, which fires on a numbered ›/❯/▶ selection cursor (› 1.) or a footer line carrying confirm/cancel/navigate hints (esc to cancel / enter to confirm / ↑/↓ to navigate) that looks like a footer — ≥2 hints, a · separator, or a short hint-only line — rather than prose quoting the phrase. It's cursor-glyph independent, so most TUIs (Codex, Gemini, Cursor) are covered with no per-CLI code. Verify: trigger a multi-option dialog in your CLI and confirm ↑/↓ navigate it. If they don't, your dialog renders differently (e.g. a hint-less footer) — override screen_shows_selection_dialog with a detector for it (keep it bottom-of-screen scoped so response prose doesn't false-match). This method is used only for the arrow filter, so a false positive is cheap (the arrow just reaches the CLI's native handling).

3. extract_last_user_prompt(...) — accurate "Last Msg". Without it the monitor's Last-Msg column falls back to Leap's PTY recently_sent capture, which can include stray echoed keystrokes (a leading 2 showed up as "2hi"). Read the last user prompt from the transcript instead (resolve via latest_transcript_for(storage_dir, cli_name or self.name, tag) — cli_name is passed by the monitor so a custom CLI atop your provider resolves its own cli_sessions subdir); return '' on any miss so the PTY fallback applies.

2. Register the Provider

Edit src/leap/cli_providers/registry.py:

from leap.cli_providers.<name> import <Name>Provider

_PROVIDERS: dict[str, CLIProvider] = {
    'claude': ClaudeProvider(),
    'codex': CodexProvider(),
    '<name>': <Name>Provider(),  # <-- Add here
}

3. Update Package Exports

Edit src/leap/cli_providers/__init__.py:

from leap.cli_providers.<name> import <Name>Provider

__all__ = [
    ...
    '<Name>Provider',
    ...
]

4. Hook Configuration

The configure_hooks() method on your provider class IS the hook configuration. The unified src/scripts/configure_hooks.py script automatically discovers all registered providers and calls their configure_hooks() method during make install, make update, and make reconfigure.

What your configure_hooks() must do:

1. Load the CLI's config file
2. Remove old Leap entries (search for "leap-hook.sh" marker)
3. Add entries that call the hook script with state arguments:
   • Stop hook: <hook_path> idle — Called when CLI finishes processing
   • Notification hooks (if supported): <hook_path> needs_permission, <hook_path> needs_input
4. Write the config back atomically — use atomic_write_json() (or atomic_write_text()) from leap.utils.atomic_write. The session-start gate reads these settings files concurrently and a non-atomic write can leave a half-truncated file mid-rewrite, which would make hooks_installed() return False and falsely block the user.

The hook script (leap-hook.sh) is shared across all CLIs. It:

• Reads LEAP_TAG and LEAP_SIGNAL_DIR env vars (set by get_spawn_env())
• Writes {"state": "<state>"} to .storage/sockets/<tag>.signal
• For idle state, also extracts the last assistant message from the transcript

If your CLI doesn't support hooks at all, you can implement a no-op configure_hooks(), but state detection will rely entirely on PTY output patterns and silence timeout, which is less reliable. You'll still need to implement hooks_installed() returning True unconditionally so the session-start gate doesn't block the user.

hooks_installed() — mirror image of configure_hooks():

The session-start gate (in leap-server.py:_enforce_hooks_installed_or_exit) calls provider.hooks_installed() before spawning the server. If it returns False, the server refuses to start and points the user at leap --reconfigure. This catches the "user installed the CLI after Leap" case (where install-time hook config was skipped because the binary wasn't on PATH yet) plus generic "user wiped their settings file" recovery.

Implementation pattern (wrap the whole body in a broad try/except):

def hooks_installed(self) -> bool:
    try:
        hook_script = self.hook_config_dir / "leap-hook.sh"
        if not hook_script.is_file():
            return False
        with open(<your settings file>, "r") as f:
            data = json.load(f)            # or tomllib.load, etc.
        # Walk your CLI's hook config defensively.  Use isinstance()
        # checks at every nesting level — a third-party tool or a
        # hand-edit could leave a valid-JSON-but-wrong-shape file
        # (e.g. ``"command": null`` or ``"hooks": "stringy"``), and
        # the `in` operator on a non-string raises TypeError.
        ...
        return False
    except Exception:
        return False

Critical rules for hooks_installed():

• Both halves must be true: hook script exists AND settings file references it. Either alone isn't enough (a stale settings file pointing at a wiped script is still broken).
• Never raise. Wrap the entire body in try: ... except Exception: return False. The gate calls hooks_installed() on the hot path of leap <tag> — a traceback there would crash the session with no useful remediation, while returning False at least fires the gate's friendly error pointing at leap --reconfigure. BaseException (KeyboardInterrupt, SystemExit) deliberately propagates.
• Lenient hook-entry check: any single entry referencing leap-hook.sh counts. Do NOT require specific events (Stop / Notification / etc.) — that would break older installs whenever new events are added to configure_hooks().
• isinstance() at every nesting level. Don't trust the JSON shape — data.get("hooks") could be a list, entry.get("command") could be None or an int. Always check before iterating or doing in checks.

Custom (user-defined) CLIs inherit hooks_installed() from their base provider via CustomCLIProvider.__getattribute__'s delegation — there's also an explicit def hooks_installed(self): return self._base.hooks_installed() on CustomCLIProvider to satisfy ABCMeta (the abstract-method check happens at class-creation time, before delegation can kick in). Custom-CLI authors don't write either method themselves; they pass base_provider=ClaudeProvider() (or one of the other four) to CustomCLIProvider.__init__ and base_type follows automatically. All custom CLIs are variants of one of the five base CLIs — this is a hard constraint of the project.

deconfigure_hooks() — symmetric counterpart of configure_hooks():

The CLIProvider base class provides a default deconfigure_hooks() that removes leap-hook.sh and leap-hook-process.py from hook_config_dir. The unified src/scripts/unconfigure_hooks.py --all script calls this during make uninstall.

If your provider writes anything into a CLI settings or config file during configure_hooks(), you must override deconfigure_hooks() to undo those writes. The override must:

1. Surgically remove only Leap's entries from each config file (preserve all other user settings)
2. Write back atomically — same atomic_write_json() / atomic_write_text() pattern as configure_hooks()
3. Call super().deconfigure_hooks() at the end to clean up the script files
4. Never raise — wrap each config-file operation in its own try: ... except Exception: pass. Uninstall must complete even if one file is missing or corrupt.

Implementation pattern:

def deconfigure_hooks(self) -> None:
    """Remove Leap's hook entries from ~/.<cli>/settings.json."""
    try:
        if SETTINGS_FILE.is_file():
            with open(SETTINGS_FILE) as f:
                data = json.load(f)
            if isinstance(data, dict):
                hooks = data.get("hooks") if isinstance(data.get("hooks"), dict) else None
                if isinstance(hooks, dict):
                    changed = False
                    for event in list(hooks.keys()):
                        entries = hooks.get(event)
                        if not isinstance(entries, list):
                            continue
                        cleaned = [
                            e for e in entries
                            if not (
                                isinstance(e, dict)
                                and "leap-hook.sh" in e.get("command", "")
                            )
                        ]
                        if len(cleaned) != len(entries):
                            hooks[event] = cleaned
                            changed = True
                    if changed:
                        atomic_write_json(SETTINGS_FILE, data)
    except Exception:
        pass
    super().deconfigure_hooks()

Copilot-style state restoration: If your configure_hooks() displaces an existing user value (e.g. a pre-existing status-line command), save the original to a sidecar file at install time and restore it in deconfigure_hooks(). See CopilotProvider.configure_hooks() / CopilotProvider.deconfigure_hooks() and the leap-statusline-chain file pattern.

5. Optional: Shell Launcher Script

Create src/scripts/<name>-leap-main.sh for a direct shortcut:

#!/bin/bash
# <Display Name> launcher — delegates to leap-main.sh with CLI preset
export LEAP_CLI="<name>"
exec "$(dirname "${BASH_SOURCE[0]}")/leap-main.sh" "$@"

Make it executable in the Makefile configure-shell target:

@chmod +x $(SCRIPTS_DIR)/<name>-leap-main.sh

6. Makefile: Hook Cleanup on Uninstall

Primary path (automatic): make uninstall calls src/scripts/unconfigure_hooks.py --all, which calls provider.deconfigure_hooks() for every registered provider. If your deconfigure_hooks() override is correct, no Makefile edit is needed for the primary cleanup.

Safety-net rm -f lines: The uninstall target also has explicit rm -f lines for every provider's hook scripts — these run after the venv is removed and act as a fallback if the Python cleanup fails. Add your provider's script files to that block:

@rm -f "$$HOME/.<cli_config_dir>/leap-hook.sh" "$$HOME/.<cli_config_dir>/leap-hook-process.py" 2>/dev/null || true

If your configure_hooks() copies additional scripts (e.g. a status-line script like Copilot's leap-copilot-statusline.py), add those to the safety-net line too.

7-10. Automatic — No Changes Needed

These are handled automatically by the abstractions:

• ASCII banner: print_banner() uses provider.display_name
• Monitor table: table_builder_mixin.py reads provider.display_name
• CLI selector: leap-select-cli.py reads from list_providers() + get_provider().display_name
• Shell flags: configure-shell-helper.sh generates LEAP_<NAME>_FLAGS from list_providers(). Hyphens in provider names are replaced with underscores (e.g. cursor-agent → LEAP_CURSOR_AGENT_FLAGS). leap-select.sh does the same conversion when reading the env var. If your provider name contains hyphens, verify both scripts produce matching variable names — a mismatch means the user's custom flags won't be picked up

11. Documentation & String References

Many files contain hardcoded provider names in docstrings, comments, error messages, and user-facing text. When adding a new provider, grep the entire codebase for existing provider names (e.g. claude, codex, Claude Code, OpenAI Codex) and update every list that enumerates providers. Common locations:

CLAUDE.md — Update:

• Description (line 3): Add new CLI to the list
• Project Structure: Add the new provider file under cli_providers/
• Key Classes table: Add the new provider class
• get_provider() row: Add the new provider name
• IDE Setup section (if the new CLI has IDE-specific config)

README.md — Update:

• Description line: Add the new CLI name
• Prerequisites: Add link to the new CLI's docs
• Features: Ensure text is generic ("the CLI") not provider-specific
• Links footer: Add link to the new CLI's docs

Source files with hardcoded provider lists (grep for 'claude', 'codex' and Claude.*Codex):

• cli_providers/__init__.py — Module docstring
• cli_providers/base.py — Docstring examples for name, command, display_name, hook_config_dir
• cli_providers/registry.py — get_provider() docstring
• server/server.py — Usage messages, LeapServer.__init__ docstring, parse_options() docstring
• server/metadata.py — SessionMetadata.__init__ docstring
• server/state_tracker.py — Module docstring
• server/pty_handler.py — Module docstring, __init__ docstring
• utils/terminal.py — print_banner() docstring example
• scripts/leap-hook.sh — Header comments (provider list + stdin format)
• scripts/leap-main.sh — Comment listing launcher scripts
• scripts/leap-select-cli.py — Error message for no CLIs found
• scripts/leap-select.sh — Comment about per-CLI env var flags

Slack integration (grep for Claude in src/leap/slack/):

• slack/bot.py — Module docstring, class docstring, comments
• slack/output_watcher.py — Module docstring, _PROVIDER_DISPLAY_NAMES dict, method docstrings
• slack/output_capture.py — Module docstring, method docstrings
• scripts/setup-slack-app.sh — Slack app description string

Other:

• __init__.py (root package) — Module docstring
• pyproject.toml — Project description
• monitor/leap_sender.py — Docstrings referencing "Claude"

Understanding State Detection

State detection is the most complex part. There are three mechanisms:

A. Hook-Based (Primary, Most Reliable)

The CLI calls hook scripts on lifecycle events. The hook writes state to a signal file. The state tracker reads this file.

• Stop hook → writes idle (CLI finished processing)
• Notification hooks → writes needs_permission or needs_input
• State tracker reads .storage/sockets/<tag>.signal each poll cycle

B. PTY Output Pattern Matching (Secondary)

The state tracker watches raw PTY output for patterns:

• trust_dialog_patterns: Startup dialog detection (before user input) → needs_permission
• dialog_patterns: Startup dialog detection (before user input, fallback) — checked for ALL patterns present → needs_permission
• interrupted_pattern: ANSI-stripped output after user input → interrupted

Note: During running state, permission detection relies solely on Notification hooks (signal file). PTY dialog_patterns are only checked at startup.

For full-screen TUIs (Ratatui), PTY output is unreliable because screen redraws produce constant output. Set output_triggers_running = False and rely on hooks.

C. Silence Timeout (Fallback)

If no output for silence_timeout seconds while in running state → transition to idle. This catches cases where hooks don't fire.

State Machine Summary

                  ┌─── hook: idle ──────────────┐
                  │                              │
                  ▼                              │
    ┌──────────┐     send()     ┌───────────┐   │
    │   IDLE   │ ──────────────▶│  RUNNING   │───┘
    │          │◀───────────────│            │
    └──────────┘  silence/hook  └───────────┘
         │                          │    │
         │  Escape                  │    │ hook: needs_*
         │  (race)                  │    │ or PTY pattern
         ▼                          │    ▼
    ┌──────────────┐                │  ┌───────────────────┐
    │ INTERRUPTED  │◀───────────────┘  │ NEEDS_PERMISSION  │
    │              │   PTY pattern     │ NEEDS_INPUT       │
    └──────────────┘                   └───────────────────┘

Real-CLI gotchas (learned adding GitHub Copilot, a hookless animated TUI)

• The idle prompt may not be quiet. Some TUIs animate their input box and emit PTY output continuously even while idle - but only in a focused real terminal; a bare pexpect PTY stays quiet, so a headless capture lies. Continuous output defeats every silence-based idle fallback (the session sticks on RUNNING) and makes cursor visibility a useless busy/idle signal. The fix is to drive idle off a stable footer string (idle_indicator_patterns), not silence. So: verify idle / interrupt behavior against a real focused terminal (or a live .storage/state_logs/<tag>.log), never just a headless pexpect run.
• Once footer-driven, _reset_screen() on a transition is a trap. If the footer-detector must later exit the state you're entering (e.g. INTERRUPTED -> idle), do NOT wipe the screen on the way in: an incremental-repaint TUI thinks the footer is already drawn and never re-emits it, so the idle footer never returns and the session sticks. Keep the screen (gate the reset off for idle_indicator_patterns providers). Resetting is only safe on a transition into idle - the resting state the detector no longer needs to leave.
• Don't assume Escape interrupts. Copilot ignores Escape mid-turn and cancels on Ctrl+C; set interrupt_key (it defaults to Escape).
• A "question" is not a "permission". If the CLI asks the user questions and asks tool-permission, map the question to needs_input (via input_dialog_patterns), not needs_permission - else ALWAYS-mode auto-approve auto-answers it for the user. Their footers can differ ("enter to confirm" vs "enter to select").
• Gate every shared state-tracker change on a provider property that defaults to current behavior. A new CLI's quirks must never alter the shared tracker unconditionally - a blanket change regressed Claude this session. Defaults must keep the other CLIs byte-identical; prove it with the full unit + integration suites before committing.

Testing Your Provider

1. Verify Registration

poetry run python -c "
from leap.cli_providers.registry import get_provider, list_providers
print(list_providers())
p = get_provider('<name>')
print(f'{p.name}: {p.display_name}, cmd={p.command}')
print(f'hook_dir={p.hook_config_dir}, requires_binary={p.requires_binary_for_hooks}')
print(f'base_type={p.base_type}, hooks_installed={p.hooks_installed()}')
"

2. Verify Hook Configuration

PYTHONPATH=src:$PYTHONPATH poetry run python src/scripts/configure_hooks.py <name> src/scripts/leap-hook.sh

After running this, provider.hooks_installed() must flip from False to True. If it doesn't, your hooks_installed() and configure_hooks() aren't symmetric — the gate at session start will block users with no recovery (running leap --reconfigure would re-run configure_hooks(), which still wouldn't satisfy hooks_installed()).

3. Run Existing Tests

poetry run pytest tests/ -v

Existing tests should still pass. Consider adding provider-specific tests to tests/test_state_tracker.py following the Codex test patterns.

4. Manual Testing

1. Start a server: leap test-<name> --cli <name>
2. Verify the ASCII banner shows the correct CLI name
3. Open the monitor — verify the CLI column shows the display name
4. Trigger state transitions and verify detection works:
   • Send a message → state should go to running
   • Wait for completion → state should go to idle
   • Trigger a permission dialog → state should go to needs_permission
   • Press Escape → state should go to interrupted

5. Write State Tracker Tests

Add a test class in tests/test_state_tracker.py:

class TestMyCliProvider:
    """Tests for MyCLI-specific state detection."""

    def test_my_cli_specific_behavior(self, tmp_path):
        from leap.cli_providers.<name> import <Name>Provider
        provider = <Name>Provider()
        t = [0.0]
        tracker = CLIStateTracker(
            signal_file=tmp_path / "test.signal",
            clock=lambda: t[0],
            provider=provider,
        )
        # Test your provider's specific behaviors...

Checklist

Core implementation

• Provider class created in src/leap/cli_providers/<name>.py
• All abstract properties and methods implemented
• Provider registered in registry.py
• Provider exported in __init__.py
• configure_hooks() installs hooks correctly and writes atomically (use leap.utils.atomic_write)
• hooks_installed() is the symmetric inverse of configure_hooks() — both halves checked, never raises, lenient on which hook events are present
• After running configure_hooks(), hooks_installed() flips to True
• deconfigure_hooks() undoes everything configure_hooks() wrote (if you write to settings files); calls super().deconfigure_hooks(); never raises; tested manually or via unit test that settings are cleanly restored
• hook_config_dir points to correct location
• requires_binary_for_hooks set correctly
• Leap Resume feature wired (if the CLI supports resume): supports_resume, extract_session_id, resume_args — or explicitly decide to skip
• Cross-cwd resume / move mechanism wired (only when supports_resume = True): - [ ] requires_cwd_bound_resume set correctly (True if the CLI's resume needs cwd to match its recorded path) - [ ] relocate_session() implemented — file-move (Claude/Gemini/Cursor pattern) or logical no-op (Codex pattern, just calls on_committed) - [ ] If file-move: created src/leap/utils/<name>_session_move.py using the shared relocation.py primitives (signals_blocked, stage_copy_*, commit_*, must_remove_tree, make_tmp_path) - [ ] session_exists() overridden if your CLI's records have empty transcript_path - [ ] Verified the Original and Current picker options both produce a working resume (manually test from a cwd different than the recorded one)

Shell & Makefile

• Shell launcher script created (src/scripts/<name>-leap-main.sh)
• Makefile: chmod +x for launcher script in configure-shell target
• Makefile safety-net rm -f block updated with any extra scripts your configure_hooks() copies (the primary cleanup runs via unconfigure_hooks.py --all → deconfigure_hooks() automatically)
• If provider name contains hyphens: verify LEAP_<NAME>_FLAGS uses underscores in both configure-shell-helper.sh and leap-select.sh

String references (grep for existing provider names!)

• cli_providers/__init__.py — module docstring
• cli_providers/base.py — docstring examples (name, command, display_name, hook_config_dir)
• cli_providers/registry.py — get_provider() docstring
• server/server.py — usage messages (grep Usage: in main()), docstrings
• server/metadata.py — docstring
• server/state_tracker.py — module docstring
• server/pty_handler.py — module docstring, __init__ docstring
• utils/terminal.py — print_banner() docstring
• scripts/leap-hook.sh — header comments (provider list AND per-CLI stdin format comment)
• scripts/leap-main.sh — comment listing launcher scripts
• scripts/leap-select-cli.py — error message
• scripts/leap-select.sh — env var flags comment
• slack/bot.py — docstrings and comments
• slack/output_watcher.py — _PROVIDER_DISPLAY_NAMES dict, docstrings
• slack/output_capture.py — docstrings
• scripts/setup-slack-app.sh — app description
• src/leap/__init__.py — package docstring
• pyproject.toml — project description
• monitor/leap_sender.py — docstrings

Documentation

• CLAUDE.md updated (description, Project Structure, Key Classes table)
• README.md updated (description, prerequisites, links footer)

Testing & verification

• Existing tests pass (poetry run pytest tests/ -v)
• Provider-specific tests added
• Manual testing: server startup, state transitions, monitor display
• Self-verification: grep -rn for old provider names to catch stragglers