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
(
logmindorclud-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.
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_tokensin 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-*.ymlorstrict-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 doctorflags 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 — when / how to log decisions; reading prior context.
- clud-bug-collaboration — 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.