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 commands/lugassawan/swe-workbench/capturegit clone --depth 1 https://github.com/lugassawan/swe-workbenchWhat 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 | $0.00022 | $0.01470 |
| Opus 5 | $0.00011 | $0.00735 |
| Sonnet 5 | $0.00004 | $0.00294 |
| Haiku 4.5 | $0.00002 | $0.00147 |
Grade A, and why
capture 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 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.
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 — 60 lines — stays where its author put it; the contents beside it link to each section on GitHub.
The user wants to capture: $ARGUMENTS
Interrogation mode. Before producing anything, resolve the mode:
- Explicit signal in the invocation is honored without asking. grill-me =
--grill, "grill me", or "grill-me mode". standard =--standard, "standard", or "quick". Strip the signal from $ARGUMENTS and record the resolved mode. - No explicit signal: ask via
AskUserQuestion— one question, header "Mode", options Standard (recommended, listed first) and Grill me. Standard description: "Lightweight clarify — a restatement and at most one question, then proceed." Grill-me description: "Relentlessly walk the decision tree one question at a time, each with a recommended answer, self-answering from the codebase where possible." Use the user's choice.
Standard mode: proceed with the command's existing lightweight clarify (a restatement and at most one clarifying question) — do not ask the mode question again.
Grill-me mode: activate swe-workbench:workflow-grill and run its interrogation loop to completion (exit on shared understanding or when the user says "proceed"). Then thread the emitted ## Resolved decisions block into the command's normal artifact/delegation step below — the same way a ticket-context summary is prepended — and continue as in standard mode.
Delegate to the swe-workbench:product-manager subagent. Its response must deliver all of the following before any issue is filed:
-
Auth + repo detection. Run
gh auth status, thengh repo view --json nameWithOwner -q '.nameWithOwner'. Surface the detected repo in the preview asFiling into: <owner>/<repo>. If either command fails, bail with a clear single-line message ("Repo detection failed: . Rungh repo viewto diagnose.") and stop. -
Restatement. One sentence in the user's domain language confirming the thought. If the thought is ambiguous, ask exactly one clarifying question before continuing.
-
Product framing — four lenses applied with brevity:
- Problem. User pain or constraint, stated as the user's pain — not the feature.
- Value. Who benefits and how. One "so-what" sentence.
- Acceptance criteria. 2–4 bullets (Given/When/Then or simple bullets).
- Impact / Effort (RICE-lite).
Impact: S/M/LandEffort: S/M/L, one sentence each.
-
Template discovery. List
.github/ISSUE_TEMPLATE/filtered to*.md, skippingconfig.yml. Read each template's frontmatter and first ~20 body lines. Classify the thought into the closest-fit template with a one-sentence reason, or note "No issue templates found; using default body shape" when none exist. -
Label discovery. Run
gh label list --json name -q '.[].name'to get the repo's available labels. If the command fails or returns empty output, treat the label list as empty and proceed directly to step d (no match → omit--label). Otherwise select a label using this chain:a. Template frontmatter: if the chosen template has a
labels:field and that value exists verbatim in the repo's label list, use it. b. Fallback — substring match (case-insensitive): if the frontmatter label is not present verbatim, pick the first repo label whose name case-insensitively contains (or is contained by) the template's value. c. No template chosen: map by commit-tag —[feat]→enhancement,[bug]→bug,[chore]→documentation— then apply the same chain against the repo's label list. If no commit-tag is recognisable in the user's input, proceed directly to step d. d. No match found: omit--label; record this so the preview can warn the user ("No matching label found; filing without label"). -
Duplicate scan.
gh issue list --search "<2-3 keywords>" --state open --limit 5. Surface matches. Ask before drafting if any look duplicative. -
Draft. With a template: fill its sections, prepend
## Product framing. Without a template: use## Problem/## Value/## Acceptance criteria/## Impact / Effort/## Additional context.
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 · 60 lines · 22 tokens per session scan A fe4595084792
capture is a command published in the GitHub repository lugassawan/swe-workbench (2 stars, last pushed 2d ago), licensed MIT. It adds 22 tokens to every session and 1,470 once invoked, about $0.0001 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 2026-08-31.
Other commands, from other repositories
brooks-audit
Run a Brooks-Lint architecture audit.
brooks-debt
Run a Brooks-Lint tech debt assessment.
brooks-health
Run a Brooks-Lint codebase health dashboard across all four dimensions.
getting-started
Interactive 10-minute guided tour for new Claude Craft users.
ralph-run
Executer Claude en boucle continue jusqu'a completion de la tache (Ralph Wiggum v2.0).
daily-standup
Génération Résumé Daily Stand-up.