cogmeta-talon 0.7.85

Context intelligence for Agentic Development — MCP server + setup CLI

Talon

Context intelligence for Agentic Development — graph-based code context for AI coding agents.

Supported Languages

Language	Extensions	Parser Features
Python	.py	Classes, decorators, raise/except, async
TypeScript	.ts, .tsx	Classes, interfaces, arrow functions, ES imports
JavaScript	.js, .jsx, .mjs, .cjs	Arrow functions, ES imports/exports, method chains, error paths
Go	.go	Structs, receiver methods, interfaces, panic/recover
Java	.java	Annotations, constructors, throws, inheritance, enums
Rust	.rs	Structs, traits, impl blocks, enums, panic!/? error paths
C#	.cs	Classes, interfaces, attributes, constructors, throw/catch
Kotlin	.kt, .kts	Data/sealed classes, companion objects, suspend functions
Ruby	.rb	Classes, modules, include/extend, singleton methods, rescue
PHP	.php	Classes, interfaces, traits, enums, namespace imports

All languages get full-depth parsing: classes/structs, functions/methods, imports, call edges, error flow edges, method chain resolution, and cross-module call edges.

Install

pip install talon

Quick start

# Authenticate with cogmeta.ai (opens browser, no copy/paste)
talon-setup auth login

# Generate .mcp.json for your project
talon-setup init --cloud

Then restart Claude Code — Talon is active.

talon work — SPECx client-pull dispatcher (new in 0.7.80)

Poll the SPECx work queue, run claimed tickets locally via claude -p, stream audit events back, post completion. Closes the loop for client-pull dispatch (Slice 2).

# Smoke-test: claim + start + audit + complete one ticket without invoking claude
talon work --once --dry-run

# Real run: claim and run one ticket via claude -p (180s hard cap)
talon work --once --timeout-seconds 180

# Long-running worker: drain the queue until interrupted
talon work --poll-seconds 5

# Bounded run: process exactly N tickets then exit
talon work --max-runs 10

Prerequisites for the real path (without --dry-run):

1. talon auth login — provisions ~/.talon/credentials.json with your API key.
2. claude /login — your local Claude Code CLI must be authenticated.
3. ~/.claude/settings.json must have permissions.defaultMode: "bypassPermissions" (or you'll see the wrapper report status=done with no files produced — see the talonpilot reproduction in postmortem-slice-2-day-2026-05-16.md).

Concurrent agents. Multiple talon work processes can race the same queue safely. Each --client-id gets its own per-second polling budget (server-side rate limit is keyed per client_id). Verified: 3 concurrent agents, 9 tickets, 173 seconds wall-clock, 6/6 pytest passing on produced code.

Sandbox first. talon work writes files into the current working directory and creates branches like specx/T-0042. Use a clean subdirectory:

mkdir specx-work && cd specx-work && git init
talon work --once
git log --all --oneline    # see what got produced

On SIGINT (Ctrl-C) / SIGTERM the wrapper POSTs /release on the active claim so the queue doesn't starve.