Deploy agent hooks: Claude PostToolUse/Stop + Cursor rules#94
Merged
Conversation
Adds an agent_hooks task that links the in-repo `claude/settings.json` into `~/.claude/` and the `cursor/rules` directory into `~/.cursor/`, so both agents pick up shared config from this dotfiles checkout. Includes the initial set of Cursor rules (planning, dbt). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two command-type hooks replace earlier prompt-type hooks that would narrate past their own stated gate (e.g. flagging a JSON config edit as "a YAML build artifact"). Command hooks make the path filter deterministic in shell — the gate either matches or the script exits silently, no model interpretation involved. - `plan-review.sh` (PostToolUse on Write|Edit): only fires when the edited file is under `.cursor/plans/` or `.claude/plans/`. On a match, injects a strict plan-quality checklist as additional context. - `stop-check.sh` (Stop): only fires when the current turn edited a `.sql` file inside a dbt project (has a `dbt_project.yml` ancestor) or any `.qmd` file, AND no build/render command ran. Suggests the matching command per file group. Both scripts are referenced via `$HOME/...` so the deployment is portable across machines and accounts. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Extracts the inline heredoc out of `plan-review.sh` into a dedicated markdown file so the prompt picks up editor/formatter support and so future prompts can sit alongside it under `claude/prompts/`. The script resolves the file relative to its own location, so it remains portable across machines. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Previously `plan-review.md` and `planning.mdc` each restated the same
five core gates (exit criteria, invariants, failure modes + premortem,
assumptions & unknowns, outside view) plus the conditional reversibility
and System 2 gates. Editing one and forgetting the other would silently
drift the contract.
The hook script now reads `cursor/rules/planning.mdc` (strips its YAML
frontmatter) and substitutes its body into the `{{PLANNING_RULE}}`
placeholder in `plan-review.md`. `plan-review.md` keeps only the audit
framing and the `PLAN OK` / gap-list output contract.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The `claude/prompts/plan-review.md` template only held ~10 lines of
audit framing + an output contract wrapping a `{{PLANNING_RULE}}`
placeholder. Claude Code has no convention for a `prompts/` directory,
and the substitution loop was more indirection than the content
deserves. Inlining the framing as two small heredocs around the
canonical `planning.mdc` body keeps the single-source-of-truth
property and drops a file.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Removes from `cursor/rules/planning.mdc`: - "Reversibility" and "System 2 triggers" conditional gates — these are geared at production deploys / data migrations rather than the analytical/dbt work that's the actual day-to-day. - "Three-tier boundaries for actions" — overly prescriptive, and the "Never do" list in particular bakes in dbt-shop-specific anti-patterns that don't generalize. Five core gates (exit criteria, invariants, failure modes + premortem, assumptions & unknowns, outside view), minimal viable change, viz confirmation, and plan format all retained. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Plan-review audit prompt now follows judge-prompt best practice: named gates, per-gate PASS/FAIL/N-A verdicts with quoted evidence, strict default on uncertainty, XML-delimited rule body, and the full rule injected every edit (session marker went stale after compaction). Stop-check nag spells out what verification means and gives an honest exit for false positives. New track-tool-use.sh records this turn's edits/commands to a state file so stop-check can target dbt/quarto files without parsing the transcript. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…sions planning.mdc: plans for new/changed dbt models must declare the grain (operator-reviewed spec, not implementation-discovered) and exit via a passing uniqueness test on it. dbt.mdc: grain test required per model, failing grain tests are findings (never widen keys or dedup to pass), all dedup must be plan-sanctioned, marts need a reconciliation test against an externally controlled number. New dbt-rules.sh hook mirrors Cursor's glob-scoped rule loading for Claude Code: injects dbt.mdc on the first .sql/.yml edit inside a dbt project per session, then a compressed reminder of the load-bearing rules on later edits. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
agent_hookstask symlinksclaude/settings.jsoninto~/.claude/andcursor/rules/into~/.cursor/, plus the initial Cursor rules (planning, dbt).plan-review.shfires only for edits under.cursor/plans/or.claude/plans/, then injects a strict plan-quality checklist.stop-check.shfires only when the current turn edited a.sqlinside a dbt project (dbt_project.ymlancestor) or any.qmd, AND no build/render command ran.$HOME/...for portability across machines/accounts.The hooks replace earlier
prompt-type variants whose sub-models narrated past their own stated gate (flagging non-matching edits anyway). Command hooks make the gate deterministic — match or silent exit.Test plan
$HOMEexpansion insh -cinvocation.100755) on both.shfiles.ansible-playbook ansible/dotfiles.yml --tags hookson a fresh box and confirm both symlinks land.plan-review.shinjects the checklist..sqledit inside a dbt project with nodbt buildand confirmstop-check.shreminds.Dependency note: scripts require
jq, which is already assumed by other dotfiles utilities (preview_hocon,grafana/query_grafana.bash) and provisioned outside this repo.🤖 Generated with Claude Code