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/alphabravo-oss/guild/helpgit clone --depth 1 https://github.com/alphabravo-oss/guildWhat 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 | $0.00014 | $0.01867 |
| Opus 5 | $0.00007 | $0.00933 |
| Sonnet 5 | $0.00003 | $0.00373 |
| Haiku 4.5 | $0.00001 | $0.00187 |
Grade A, and why
help 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 — 142 lines — stays where its author put it; the contents beside it link to each section on GitHub.
The user wants help with the adhoc plugin. Reply with exactly this message — verbatim, no additions, no embellishment:
adhoc — methodical-mode for ad-hoc Claude work (v0.3.0)
WHAT IT DOES
Always-on UserPromptSubmit hook prepends a methodical pre-response checklist
to every user turn: restate the task, mark assumptions VERIFIED/UNVERIFIED,
Read Floor (read code before editing), Approach Deliberation (>=2 candidates),
Blast Radius (grep callers), Stall Check (anti-spin), Competing Hypotheses
(for bugs), Goal-Driven Execution (verifiable success criteria), Restraint
Check (no over-engineering). Stops the race-to-an-answer pattern.
Always-on Stop hook enforces four gates after the response is drafted:
(a) Citations — file:line citations must be backed by Read/Grep this turn
(b) Uncertainty — blocks "not verified" / "haven't checked" / "I assumed"
(c) Grounding — codebase questions require a Read/Grep/Glob this turn
(no general-knowledge fallback)
(d) Iterative Critic — substantive responses must pass a multi-round
critic dialog (NEW in v0.3.0):
- Rounds 1-2: adhoc:fast-critic (Haiku, cheap, 10-item rule audit)
- Rounds 3-5: adhoc:pre-stop-critic (opus escalation, same audit
with deeper file-reading verification)
- Each critic sees prior-round feedback and checks if Claude
ADDRESSED, STILL FAILING, or REPHRASED-NOT-FIXED each prior flag
- At round 5 without CONCUR: HARD block — Claude cannot Stop until
/adhoc:trust-me bypass is invoked by the user
Gates (a)/(b)/(c) honor stop_hook_active (one block per turn each).
Gate (d) uses a transcript-based round counter (up to 5 block-rounds).
COMMANDS — methodical-mode (UserPromptSubmit hook)
/adhoc:status Show current methodical-mode state
/adhoc:on Re-enable methodical-mode for this session
/adhoc:off Disable methodical-mode for the rest of this session
/adhoc:casual Skip methodical-mode for the NEXT turn only — auto-reverts
/adhoc:deep Heavier methodical analysis pass (skill, read-only)
/adhoc:help This message
COMMANDS — Stop-hook gates
/adhoc:citations-on Enable file:line citation verifier (default)
/adhoc:citations-off Disable citation verifier
/adhoc:uncertainty-on Enable uncertainty-tell scanner (default)
/adhoc:uncertainty-off Disable uncertainty-tell scanner
/adhoc:strict-on Enable grounding audit + iterative critic gate (default)
/adhoc:strict-off Disable grounding audit + iterative critic gate
/adhoc:trust-me One-shot bypass of grounding + critic gates for the
NEXT turn only — auto-consumed. Uncertainty + citation
gates stay on. Use this to escape a HARD-blocked
round-5 deadlock when the critic is misbehaving.
SUBAGENTS
adhoc:second-opinion Independent fresh-context critique of a plan / approach.
User-invoked. Heavier review for major decisions.
Not part of the Stop-hook gate dialog.
adhoc:fast-critic Hook-invoked Haiku critic — runs rounds 1-2 of the
gate dialog. Cheap 10-item rule audit per round.
Emits sentinel [adhoc:fast-critic-output] so the Stop
hook skips itself on critic output (recursion guard).
adhoc:pre-stop-critic Hook-invoked opus critic — runs rounds 3-5 of the
gate dialog (escalation tier). Same 10-item audit
with deeper file-reading verification and
prior-round continuity check. Emits sentinel
[adhoc:pre-stop-critic-output] for recursion guard.
ITERATIVE GATE DIALOG (gate d — the new v0.3.0 layer)
Round 1: Claude drafts -> Stop hook spawns fast-critic
- CONCUR -> response lands
- Non-CONCUR -> Claude revises based on specific feedback
Round 2: Claude retries Stop -> hook spawns fast-critic again with
prior-round feedback included
- Critic checks each prior flag: ADDRESSED | STILL FAILING |
REPHRASED-NOT-FIXED
- CONCUR -> response lands
Round 3: Escalates to pre-stop-critic (opus). Same loop with deeper review.
Round 4: Pre-stop-critic continues iteration.
Round 5: Pre-stop-critic last chance. If still non-CONCUR -> HARD block.
User must invoke /adhoc:trust-me to bypass.
Cost profile:
Typical turn (CONCUR round 1): 1 Haiku call (~$0.001)
Stuck turn worst case: 2 Haiku + 3 opus calls (~$0.10-0.15)
Deadlock risk: if a critic misbehaves (returns false PUSH BACK loops),
the session locks up at round 5 until /adhoc:trust-me is invoked.
This is intentional per the strict-on configuration.
STATE FILES (all under ~/.claude/)
.adhoc-state methodical-mode injector: absent/"on" / "off" / "casual"
.adhoc-citations-mode gate (a): absent/"default" / "off"
.adhoc-uncertainty-mode gate (b): absent/"default" / "off"
.adhoc-strict-mode gates (c) and (d): absent/"default" / "off"
.adhoc-trust-me one-shot bypass file for (c)+(d) — consumed on read
.adhoc-citations-log.jsonl per-turn gate decisions (audit log, includes
rounds_completed and prior_verdicts for gate d)
WHEN TO RELEASE-VALVE
/adhoc:casual For trivial chitchat where the methodical preamble is pure
noise. Skips the UserPromptSubmit injection for one turn.
/adhoc:trust-me For genuine general-knowledge questions OR when the critic
gate gets stuck in a false-PUSH-BACK loop. Skips grounding
gate (c) + iterative critic gate (d) for one turn. Citation
+ uncertainty gates still run.
/adhoc:off For long pairing sessions where you've calibrated and the
preamble repeats itself. Silences the UserPromptSubmit
injection for the rest of the session. Stop gates keep firing.
/adhoc:strict-off + /adhoc:uncertainty-off + /adhoc:citations-off
Nuclear: disables ALL Stop gates for the session.
DESIGN NOTES
- Default is always-on enforcement. If you remembered to invoke it, you
wouldn't need it — the point is that it fires when you forget.
- Citation, uncertainty, and grounding gates use deterministic regex.
Critic gate (d) uses LLM judgment (Haiku then opus) because language
variation is too rich for regex (e.g., "haven't checked" vs "re-checking"
vs "let me look into it" — all the same dodge, different vocabulary).
- Recursion guard: critic responses begin with their sentinel marker.
Stop hook detects EITHER fast-critic or pre-stop-critic sentinel and
skips ALL gates on those responses.
- All gate decisions are logged to .adhoc-citations-log.jsonl for
post-hoc auditing of false positives / negatives.
- The 10-item compliance audit each critic runs:
1. Rules consulted (CLAUDE.md/AGENTS.md/memory)
2. Read Floor satisfied
3. Comments-not-code (no docstring citations passed off as evidence)
4. Approach Deliberation written
5. Blast Radius written
6. Competing Hypotheses written (for bugs)
7. Restraint check (no scope creep)
8. Self-critique
9. Promise-without-action
10. Hedge-laundering
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 · 142 lines · 14 tokens per session scan A 1cdf757488e2
help is a command published in the GitHub repository alphabravo-oss/guild (2 stars, last pushed 2d ago), licensed MIT. It adds 14 tokens to every session and 1,867 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
git
Git operations with intelligent commit messages and workflow optimization.
checklist
Generate a custom checklist for the current feature based on user requirements.
clarify
Identify underspecified areas in the current feature spec by asking up to 5 highly targeted clarification questions and encoding answers back into the spec.
specify
Create or update the feature specification from a natural language feature description.
analyze
Perform a non-destructive cross-artifact consistency and quality analysis across spec.md, plan.md, and tasks.md after task generation.
constitution
Create or update the project constitution from interactive or provided principle inputs.