Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/marcus/td/agents-mdgit clone --depth 1 https://github.com/marcus/tdWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.01104 | $0.01104 |
| Opus 5 | $0.00552 | $0.00552 |
| Sonnet 5 | $0.00221 | $0.00221 |
| Haiku 4.5 | $0.00110 | $0.00110 |
Grade C, and why
td AGENTS.md scanned grade C with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Hidden instructionshighPrompt injection
Directives inside HTML comments, invisible characters or bidirectional overrides are read by the model and not by the person reviewing the file.
<!-- td-agent-instructions:start --> How it starts
The opening of the file, as written. The whole thing — 107 lines — stays where its author put it; the contents beside it link to each section on GitHub.
AGENTS.md
Working with td
td keeps task context durable across sessions. In a new context, run td usage --new-session -q to see current work.
Use your judgment about how much tracking a task needs. For substantive work: td start <id>, record progress with td log, hand off with td handoff <id>, then td review <id>.
Closing needs a review. Say who did it (default trusted mode; delegated/strict allow only the first):
- independent session:
td approve <id> --reason "..." - a sub-agent:
td approve <id> --reviewed-by "<who>" - you:
td approve <id> --self-review --reason "..."
Prefer a reviewer with its own TD_CONTEXT_ID; never name one who did not review.
Run td usage or td <command> --help.
Review Model (Trusted Review)
Work needs a review before it closes, and td asks who performed it. In the
default trusted mode an independent session approves with td approve <id> --reason "..."; a reviewer can instead attest without closing via
--record-only and any session closes after. When you implemented the work
yourself, name the reviewer with --reviewed-by "<who>", or acknowledge a
genuine self-review with --self-review --reason "...".
--reviewed-by is an attestation td cannot verify. Naming a reviewer who did
not review is worse than an honest self-review, because it reads as independent
in the audit trail. Do not create a throwaway session to make a review appear
independent either. Pin review_policy_mode=delegated|strict when a project
needs a mechanical independence boundary rather than an honesty-based one.
Development Approach
- Be pragmatic:
tdis a focused local tool, not an enterprise platform. - Use worktrees for major features, risky migrations, or parallel work. Make quick localized fixes on the current branch when scope and verification are clear.
- One independent review and one rejection cycle is normally enough. Continue only for a genuine P0 data-loss/security finding; track other observations as follow-up work.
- Test likely failures and important boundaries, not exhaustive hypothetical states without a concrete use case.
- Surface friction after the first surprising blocker or material scope growth. Pause and explain it rather than silently turning a small fix into a subsystem.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 107 lines · 1,104 tokens per session scan C 163d061662dd
td AGENTS.md is an instructions file published in the GitHub repository marcus/td (243 stars, last pushed 5d ago), licensed MIT. It adds 1,104 tokens to every session, about $0.0055 per session on Opus 5. A static security scan graded it C with 1 finding (hidden instructions). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other instructions, from other repositories
langgraph AGENTS.md
Instructions for langchain-ai/langgraph, covering agents instructions, corridor security analysis, libraries and dependency map.
agentgateway copilot-instructions.md
Instructions for agentgateway/agentgateway: Do not check for, speculate about, or report compilation errors during code review. Compilation diagnostics from review are frequently incorrect; rely on CI to detect and report compilation failures.
hatch3r CLAUDE.md
Instructions for hatch3r/hatch3r, covering hatch3r — development instructions, architecture, development commands, two-axis pillar framework (2.0.0) and orchestrator self-discipline (bypass protection).
aeon CLAUDE.md
Instructions for aeonfun/aeon, covering aeon, how aeon works, strategy, voice and soul file hierarchy (read in this order).
ken CLAUDE.md
Instructions for townsendmerino/ken, covering claude.md, what this is, repository ownership (read this first), commands and embedding parity & golden fixtures (now in aikit).
gangsta GEMINI.md
Instructions for kucherenko/gangsta, a project described as: AI agentic skills framework for spec-driven development, built on the organizational model of mafia.