CLAUDE.md · refactoringhq/tolaria
CLAUDE.md from refactoringhq/tolaria: 107 lines covering Project Memory, Quality gate: accepted risks, Post-task reflection, Housekeeping Turns.
Template by refactoringhq
Install
Works with Claude Code. Read the repository before you give any tool access to your data.
curl -fsSL https://raw.githubusercontent.com/refactoringhq/tolaria/main/CLAUDE.md -o CLAUDE.md
Read it and adapt it to your project before use.
Get the weekly diff
The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.
About
Use cases: code-review, code-generation
Reviews
Write a reviewNo reviews yet. Installed it? Reviews come from verified installers only. How reviews work
What's in it
A CLAUDE.md file: project memory Claude Code loads at the start of every session.
- For
- Claude Code
- Length
- 107 lines
- File
CLAUDE.md
Sections
- Project Memory
- Quality gate: accepted risks
- Post-task reflection
- Housekeeping Turns
First lines (shown under the repo's AGPL-3.0 license)
--- type: Note _organized: true --- @AGENTS.md This file is only a Claude Code compatibility shim. Keep shared agent instructions in `AGENTS.md`. <!-- verity-memory:start --> ## Project Memory This project has a knowledge graph maintained at `.verity/memory/`. Before starting non-trivial work, scan `.verity/memory/index.md` for decisions, gotchas, and patterns that may apply to the change you are about to make. Open specific node files via the Read tool when the title or scope suggests relevance. The graph is auto-maintained by Verity. Files at `.verity/memory/_archive/` are superseded — ignore them unless investigating history. Knowledge the organization's other repositories learned is mirrored outside the repo at `~/.verity/orgs/<host>/<owner>/memory/` (`verity memory org` lists it). Each claim says where it came from; one that is wrong here is demoted for everyone with `verity memory demote <id> --reason "…"`. ## Quality gate: accepted risks When the Verity pre-commit/pre-push gate FAILs, fix the findings — that is the default. Use `verity waive <pattern-id> --file <path> --reason "…"` ONLY to relay a risk a human has explicitly accepted: a named code-review finding, an ADR, or the user saying so in this conversation. The --reason must cite that source. Never waive on your own judgment, to get past a block, or pre-emptively. A waive binds to the file's current bytes and voids automatically when the file changes, and every waive is recorded in the run ledger. For a pattern-level false positive use `verity feedback finding <run-id> <pattern-id> false_positive` instead. ## Post-task reflection When a task is complete (you've created a PR, the user says "done" or "ship it", or the work is clearly finished), **draft the reflection yourself** — 1–3 concrete things worth remembering (a decision, a gotcha, or a pattern), each cited with the files / PR / commands / error-signatures it came from. Skip entirely if nothing non-obvious happened — that judgement is the ONLY filter, because nothing reviews the reflection before it lands. Then record it straight away. There is no confirm step, in any environment: ```bash verity reflect --user-input - --kind <kind> <<'VERITY_EOF_<random>' <your draft> VERITY_EOF_<random> ``` The draft goes through the quoted heredoc, never inside `"…"` on the command line, where a `"` or `$(…)` in it would run in the shell. Replace `<random>` with 8 random letters and digits that you pick for this call, the same at both ends. Text written before you picked them cannot end the heredoc early. Add `--confirmed` ONLY when the user authored or dictated the words. Without it the node is stored as `source: agent` — Verity thought this, nobody checked it. With it, `source: user` at full confidence — a person stands behind this. Never claim the second for your own draft, however good it is. **Name the files in the text.** Verity scopes the reflection to the paths it cites, and a reflection that names no file in this repo is never retrieved for a later review — it is recorded and then invisible. The command says so when it happens; `--file-globs '<path or glob>'` is the fix when the prose cannot carry the paths. Then tell the user, in one line, what you recorded and where: the command prints the node id, the path under `.verity/memory/`, and a dashboard link. They did not agree to it in advance, so say it happened — editing or deleting that file is how they correct it. > Durable, hand-curated guidance goes in the preserve region below (it survives > regeneration) or anywhere OUTSIDE these markers. Everything else between the > markers is tool-owned and overwritten on each run. ## Housekeeping Turns When a turn will be pure housekeeping — pulling, installing dependencies, rebasing, a formatting sweep you are not authoring — declare it BEFORE doing it: ```bash verity ignore --turn --agent --reason "pulling latest before starting" ``` This skips the review for that turn, which saves the turn Verity would otherwise spend saying it had nothing to say. Use `--for 30m` instead of `--turn` when a single piece of housekeeping spans several turns. **It is a claim about the turn, not a way to silence review.** The declaration is checked against what the turn actually did: if anything is authored — by you, by a subagent, or by a shell command that can write files — it voids, the review runs anyway, and the broken declaration is reported. So declare housekeeping you are about to do, never work you have already done, and never as a way to get past a finding. Declarations are budgeted per session and every one is recorded with its reason. <!-- verity-memory:preserve --> <!-- Add binding, hand-curated guidance here; it survives Verity regeneration. --> Verity is set up only on some machines (currently the agent-claude box). If the `verity` CLI or `.verity/memory/` is missing, skip the Verity steps above; `AGENTS.md` stays the source of truth. <!-- /verity-memory:preserve --> <!-- verity-memory:end -->
Signals
- Stars
- 20,221
- Forks
- 1,389
- Open issues
- 13
- Last commit (this folder)
- Oct 8, 2026 (20 days ago)
- Last release
- Oct 13, 2026 (15 days ago)
From GitHub (fetched Oct 24, 2026).
Details
- License
- AGPL-3.0
- Repository
- github.com/refactoringhq/tolaria
- First seen
- Oct 24, 2026
- Last checked
- Oct 24, 2026
- Source
- GitHubgithub-graphql
More from refactoringhq
Cite this page
CLAUDE.md · refactoringhq/tolaria. Agent Mart. https://agentmart.dev/templates/refactoringhq-tolaria-claude-md (last updated 2026-10-24)