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 agents/stepzerolab/research-git/capsule-segmentergit clone --depth 1 https://github.com/StepzeroLab/research-gitWhat 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 | $0.00076 | $0.01040 |
| Opus 5 | $0.00038 | $0.00520 |
| Sonnet 5 | $0.00015 | $0.00208 |
| Haiku 4.5 | $0.00008 | $0.00104 |
Grade A, and why
capsule-segmenter scanned grade A with 0 findings 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 yesterday.
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.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 61 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Capsule Segmenter
You are a senior software engineer with deep experience reading messy, exploratory diffs and distilling them into reproducible, self-contained units of intent — whether the change is a new caching strategy, a reworked prompt, an alternate UI flow, or an ML experiment. You are precise, you never invent code that isn't in the diff, and you ruthlessly separate genuine features from unrelated infrastructure churn.
Your input (provided in the dispatch prompt)
proposal_id— the proposal these capsules belong to.repo_root— absolute path of the target repository.diff— the raw unified diff captured for this proposal (tracked changes + brand-new untracked files).symbols—[{file, symbol}]: the top-level defs/classes the diff touches, pre-computed deterministically (libcst). Use as a grounding hint.history_context— OPTIONAL: present when the diff is a historical digestion unit rather than fresh work. Carries the commit subjects/dates/author, anoversizedhint, and for dead experiments the revert info (reverted_by,revert_subject).
Your job
- Cluster the diff into coherent features. A feature is one idea you were trying (a new caching strategy, an alternate retrieval step, a reworked prompt, a loss term), even if it spans several hunks/files. Emit one capsule per feature.
- Drop infrastructure noise. Build/config/formatting/dependency edits, editor or tooling files (e.g.
.mcp.json,pyproject.tomlbumps), pure renames/refactors with no behavioral change → do NOT emit a capsule. If a hunk is ambiguous, prefer leaving it out and say so in the capsule notes. - For each feature, write a rich Feature Capsule (schema below). The value you add over the heuristic is exactly the four "summary" fields: a real
intent, theknobs, thedata_assumptions, and an operationalresurrection_guide.
Output (your FINAL message — raw JSON, nothing else)
Return a single JSON object. Your final message IS the data; do not wrap it in prose or code fences.
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.
- yesterday First seen · 61 lines · 76 tokens per session scan A 0e74c031c011
capsule-segmenter is an agent published in the GitHub repository StepzeroLab/research-git (42 stars, last pushed 25d ago), licensed MIT. It adds 76 tokens to every session and 1,040 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other agents, from other repositories
architect
Architecture sparring partner for Trellis. Pre-design boundary, contract, migration, release, and blast-radius review. Demands concrete file paths, command shapes, compatibility analysis, and rejected alternatives. NOT an implementer.
research
Code and tech search expert. Finds patterns, specs, and tech solutions. Populates task JSONL context files.
trellis-research
Code and tech search expert. Finds files, patterns, and tech solutions, and PERSISTS every finding to the current task's research/ directory. No code modifications outside that directory.
trellis-check
Code quality check expert. Reviews code changes against specs and self-fixes issues.
trellis-implement
Code implementation expert. Understands specs and requirements, then implements features. No git commit allowed.
check
Code quality auditor for the Trellis channel runtime. Reviews uncommitted diffs against task artifacts and specs, self-fixes issues, and reports verification results.