# Token Frugal Tooling

> Use in any project that has BOTH logmind (`.logmind/config.yml` or `docs/decisions.md`) AND clud-bug (`.github/workflows/clud-bug-review.yml`) installed. Quick-reference for the org's token-frugal conventions: env vars, artifact defaults, and clud-bug's "agent mode" flag. Detail lives in the per-tool skills — load `logmind` and `clud-bug-collaboration` alongside this one when you actually need to invoke them.

- Skill: `thrillmade/token-frugal-tooling` (Agent Skill)
- Install (CLI): `npx get-skill add thrillmade/token-frugal-tooling`
- Raw SKILL.md: https://api.skillmd.com/api/skills/thrillmade/token-frugal-tooling/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: thrillmade (https://skillmd.com/u/thrillmade)
- Updated: 2026-11-14
- Page: https://skillmd.com/skills/thrillmade/token-frugal-tooling

---


# Token-frugal tooling — quick reference

This project uses logmind + clud-bug, both engineered for token frugality.
The defaults below land automatically; this skill is just so you don't
re-learn them every session.

## When to use

- Both logmind and clud-bug are installed (see frontmatter) and you're about
  to set an agent-mode env var, read a compact on-disk artifact, or triage a
  clud-bug review comment, and want the one-line default rather than
  re-reading either tool's full skill.

## When NOT to use

- Only one of the two tools is installed — load that tool's own skill
  (`logmind` or `clud-bug-collaboration`) directly; the defaults here assume
  both are present.
- You need the mechanism behind a default, not just the default itself —
  that detail is owned by the per-tool skill this one cross-references, not
  duplicated here.

## Agent-mode env vars (set once per command or session)

| Var | Tool | Effect |
|---|---|---|
| `CLUD_BUG_QUIET=1` | clud-bug ≥ v0.6.7 | Quiet mode for the clud-bug CLI: one `ok <key-value>` line per command (`init` / `update` / `add` / `refresh` / `edit-workflow` / `review-mode`). `--quiet` / `-q` is equivalent. |

Prefer it for agent invocations even when chaining commands — every
command emits a single parseable line, so the next step can grep/parse
without re-running. logmind v1.2.0 has no quiet mode.

```bash
CLUD_BUG_QUIET=1 clud-bug update
# → ok updated: @v0.6.11, N changed, M unchanged
```

## On-disk artifacts default to brief

When you read these files, expect compact format — the legacy verbose
view is opt-in. Both apply org-wide on next regen after a logmind upgrade.

| File | Default since | Shape |
|---|---|---|
| `docs/file-structure.md` | logmind v0.5.0 | Tree capped at **depth 2**. Override per-invocation: `logmind file-structure --max-depth N` or `logmind tree --max-depth 0` for unbounded. |
| `docs/timeline.md` | logmind v0.5.4 | Per month: header with `(N decisions)`, newest entry + `... N-2 more decisions ...` elision + oldest entry. ≤2-entry months show all entries. `logmind timeline --full` recovers the legacy per-decision listing. |

## Reading clud-bug review comments (v0.6.5+)

Every Clud Bug summary comment leads with `Found: N 🔴 / N 🟡 / N 🟣` —
read this line first to triage.

- 🔴 important — bug / security / performance / missing test coverage
- 🟡 nit — style / naming / micro-optimization (advisory)
- 🟣 pre-existing — issue that pre-dates this PR; don't fix here unless asked

Per-finding shape:

```
🔴 [skill-name]: One-line claim anchored to file:line.
<details><summary>Reasoning</summary>

…

</details>
```

On a re-read, you can usually trust the headline and skip the collapsed
`Reasoning`. Expand only when the headline is ambiguous or you want the
evidence trail.

If you see a `### Diagnostics` block before the Skills-referenced footer
(clud-bug v0.6.18+ / 0.0.T): a `head -c` cap fired during the review.
The block names which section truncated and the recovery outcome (e.g.
"recovered with 2x cap", "still truncated"). Treat it as confidence
calibration — if a finding's section was deferred, the verdict on that
section is partial.

## What you don't need to do

- **Don't override clud-bug budgets** (`MAX_DIFF_BYTES`, etc.) unless
  you have a documented reason. Defaults cover 95%+ of real PRs.
- **Don't worry about caching** — clud-bug v0.6.3+ writes its stable
  prefix into the Claude Code CLI's auto-cached system layer. Cached
  input is billed at 10% of standard. Visible via
  `cache_read_input_tokens` in the run JSON.
- **Don't worry about the model** — pinned to `claude-sonnet-4-6`
  (v0.6.11). PR review fits Sonnet's profile; Opus would be ~5× the
  cost per the Anthropic pricing table.
- **Don't worry about fix-push diff cost** — clud-bug v0.6.10 reads
  only the delta since its prior review (handshake via
  `<!-- last-reviewed-sha: <sha> -->` marker in the prior summary
  comment).
- **Don't worry about workflow-only PRs** (clud-bug v0.6.14 / 0.0.W) —
  PRs touching only `clud-bug-*.yml` or `strict-mode-gate/**` skip the
  LLM review entirely. The composite-pin lock-step propagation PRs
  that used to require admin-bypass are now normal merges.
- **Don't worry about dep-bump PRs** (clud-bug v0.6.15 / 0.0.R) —
  dependabot/renovate authors AND small-diff dep-manifest-only PRs
  route to Haiku ($0.80/MTok vs Sonnet's $3) for another ~75%
  reduction on that class.
- **Don't worry about review output length** (clud-bug v0.6.16 / 0.0.X)
  — the prompt has a brevity directive ("Keep total output under
  ~600 tokens"). Per-finding 1-sentence claim + collapsible reasoning
  is the shape; you don't need to ask for short output.
- **Don't worry about prompt drift** (clud-bug v0.6.17 / 0.0.E) — the
  CI golden gate (`test/golden/{must-contain,must-not-contain,byte-budget}.json`)
  fails any PR that drops a load-bearing prompt instruction or
  re-introduces filler patterns. The byte-budget cap is intentionally
  low post-0.0.P (v0.6.20).

## Verification

- `logmind doctor` flags stale logmind templates, hooks and merge drivers
  and exits non-zero on drift. It does not probe clud-bug, and its version
  line is unreliable in v1.2.0 — check clud-bug's version with its own CLI.

## Cross-references

- [logmind](../logmind/SKILL.md) — when / how to log decisions; reading
  prior context.
- [clud-bug-collaboration](../clud-bug-collaboration/SKILL.md) — fix-push
  flow, strict mode, modifying skills, editing the workflow, full comment
  format, full cost-control wiring.

## Sources

- Version numbers and defaults above are drawn from logmind's and clud-bug's
  own release notes; the primary citation for each lives in the two skills
  above, which this page compresses.

