Index / tool

Taskmaster

Category: AI Agents
Rank: No. 1120Tools index
Pricing: Open Source
Platform: cli
Type: TOOL
Builder: @blader
GitHub: 515 stars
Added: Feb 20, 2026

About

A completion guard for coding agents that prevents AI assistants from stopping work prematurely by requiring explicit completion tokens before allowing tasks to end. Ensures agents fully complete user requests rather than stopping after partial progress with convincing summaries.

What it does

Taskmaster sits around an agent session and demands a session-specific finish marker. For interactive Codex, it watches session events and feeds a corrective prompt back through the same terminal process. For Claude, a stop hook checks the transcript and rejects an unconfirmed exit.

Why it's ranked here

The design is narrow but concrete: one shared completion contract works across two agent environments, and automation gets a deterministic signal. Same-session continuation is especially useful because it preserves working context. The value depends on accepting aggressive enforcement and a terminal-oriented Codex integration.

What's good

It checks assistant text rather than trusting a progress summary. Codex continuation stays in the running session, deduplicates completion events, and keeps monitoring after a successful turn. Claude checks the latest assistant message first, then recent transcript output. Installation and removal support either agent or both.

Tradeoffs

Codex enforcement covers interactive sessions, while recognized non-interactive command families pass directly to the real binary. Its terminal bridge requires Expect, uses fixed polling and submission timing, and offers no monitor-only mode. Claude ignores transcripts shorter than twenty lines. Unlimited blocking is the default, so missing tokens or mistaken completion judgments can prolong a session.

How to use it well

Use it for long interactive coding tasks where premature summaries are a recurring failure and downstream automation needs an unambiguous finish signal. Keep the warning cap available as an escape hatch. It does not replace tests, acceptance criteria, task planning, or independent verification of the agent’s output.

Technical notes+

run-taskmaster-codex.sh enables Codex JSONL session recording, supervises hooks/inject-continue-codex.sh, and launches the real binary through hooks/run-codex-expect-bridge.exp. The injector handles task_complete and turn_complete, deduplicates by turn ID or message checksum, and emits queued prompt files. The Expect bridge polls every 200 milliseconds, sends bracketed paste, waits 180 milliseconds, then submits. check-completion.sh implements the Claude Stop decision with jq, a per-session counter, and a search over the last 400 transcript lines. taskmaster-compliance-prompt.sh supplies both paths. The duplicate hooks/check-completion.sh differs slightly in error handling, while tests/install.test.sh references shell configuration behavior absent from the supplied current install.sh and hardcodes an author-local source path.

Observed

License: MIT
Primary language: Bash shell scripts, plus an Expect bridge and embedded Python installer logic
Install surface: Shell installer with auto-detection and explicit Codex, Claude, or combined targets
Interfaces: Codex CLI wrapper and shim; Claude Code Stop hook
Runtime requirements: Bash and jq; Codex also requires Expect, while Claude installation updates require Python 3
Supported agent environments: Codex CLI and Claude Code with Stop hooks enabled
Tests: The supplied repository tree contains shell tests for installer behavior and Codex runner routing

Read from README.md, docs/SPEC.md, LICENSE, SKILL.md, install.sh, uninstall.sh, check-completion.sh, run-taskmaster-codex.sh, taskmaster-compliance-prompt.sh, tests/install.test.sh, hooks/check-completion.sh, hooks/inject-continue-codex.sh, hooks/run-codex-expect-bridge.exp, tests/run-taskmaster-codex.test.sh, .backup-20260306T073612Z/install.sh.

What it can do

Monitor AI coding agent task completion status
AI agent work session and progress → Completion validation status
Block premature task termination
AI agent attempting to end task without completion token → Task continuation enforcement
Require explicit completion token emission
AI agent task execution → Mandatory completion token before task end
Validate full user request completion
User goals and AI agent progress → Completion verification report
Prevent false completion with partial summaries
AI agent summary and actual work done → True completion status assessment

Comments (0)

No comments yet

Indexed by a proprietary survey. Corrections welcome.