Five guardrails, from one line of config to a closed loop

This project runs entirely inside Docker. An agent that types php artisan migrate runs it on the host, which has no PHP extension set and no database. Then, being helpful, the agent invents a workaround. The failure is not the wrong command; it is that the wrong command was available.

The three artifacts

**Always use `make` targets.** Never run `npm`, `npx`, `php`, `php artisan`,
`composer`, `vite`, or similar tools directly — they execute on the host, not
in the container, and bypass the project's environment.

- No appropriate target exists → add one to the `Makefile` and document it in
  `.claude/features/common-commands.md`.
- An existing target is not flexible enough → add `ARGS=` support rather than
  reaching for the raw tool.
- Coreutils (`ls`, `grep`, `tail`, `git`, etc.) are fine to run directly.

"permissions": {
  "allow": [
    "Bash(make *)", "Bash(git *)", "Bash(gh *)", "Bash(jira *)",
    "Bash(docker compose exec *)", "Bash(grep *)", "Bash(find *)"
  ],
  "deny": [
    "Bash(php *)",      "Bash(php artisan *)",
    "Bash(npm *)",      "Bash(npx *)",
    "Bash(composer *)", "Bash(./vendor/bin/*)"
  ]
}

test-parallel:  ## Run use-case-tier tests in parallel via paratest (no DB)
test-adapter:   ## Run adapter-tier tests against SQLite :memory: / stubbed HTTP
pint-fix:       ## Fix PHP code style (ARGS= for extra flags)
migrate:        ## Run database migrations (ARGS= for extra flags)

How it fires

1. Session start toolchain.md auto-loads: its paths: glob is **, so it is in context every turn.
2. Tool call Agent tries php artisan migrate. The deny pattern matches; the call never runs.
3. Recovery The rule already named the alternative, so the retry is make migrate, not an invented workaround.

The rule's frontmatter carries the incident that motivated it. That why: field is not documentation courtesy: it lets an agent reason about the edge case the rule's letter does not cover.

---
description: Functions and patterns that must never appear in committed code —
  debug output, raw queries, imperative array functions, lazy-loaded relationships
paths:
  - "app/**/*.php"
  - "database/**/*.php"
  - "tests/**/*.php"
  - "routes/**/*.php"
layer: project
locked: true
why: A `dd()` shipped to production swallowed a KeyboardInterrupt and broke the
  dashboard; `DB::raw()` with string interpolation is a SQL-injection class; an
  N+1 in a hot path silently slows production until it pages someone.
---

# Extract NEW content from tool input — the gate judges the change,
# not the file it lands in.
case "$tool" in
  Write)     content=$(jq -r '.tool_input.content')                       ;;
  Edit)      content=$(jq -r '.tool_input.new_string')                    ;;
  MultiEdit) content=$(jq -r '[.tool_input.edits[]?.new_string]|join("\n")') ;;
esac

declare -a banned=(
  '\bdd\s*\('        '\bdump\s*\('       '\bddd\s*\('
  'DB::raw\s*\('     'DB::statement\s*\(' '\bprint_r\s*\('
  '\bvar_dump\s*\('  '\bray\s*\('
)

while IFS= read -r line; do
  case "$line" in *"charter:allow"*) continue ;; esac
  ...
done <<< "$content"

if [ -n "$violation" ]; then
  cat >&2 <<EOF
charter: blocked $tool on $rel — banned token in new content.
$violation
See .claude/rules/prohibited-patterns.md. Use the structured logger
(\`Log::debug(...)\`) for debug output and Eloquent or parameter-bound
queries instead of raw DB calls. To override one line, append the
inline marker:  // charter:allow ray
EOF
  exit 2
fi

Four design decisions worth stealing

1. Fail open No file path, non-code extension, or an allowlisted directory (scripts/, vendor/, database/migrations/) → exit 0. A sensor that blocks on cases it does not understand is a sensor you will end up turning off.
2. Exit 2 talks Exit 2 returns stderr to the agent, not to a log read after the fact. So write the block message as a prompt: what fired, why, and the specific alternative you want used instead.
3. Two escape grains Per-line // charter:allow and whole-run CHARTER_BYPASS=1. One legitimate exception should not require disabling the whole gate, because that habit is how gates die.
4. Names its rule The message cites prohibited-patterns.md. Sensor and guide stay coupled, so fixing one prompts fixing the other.

The timeline

1. T0 · Session start session-start.sh prints branch, last commit, dirty count, then confirms the git pre-commit hook exists and is executable. A missing hook is not an inconvenience; it is the silent absence of verification, so it warns loudly.
2. T1 · Before Bash check-commit-msg.sh intercepts git commit -m and tests the subject against Conventional Commits with an optional [PROJ-123] prefix. Blocks before the commit object exists: nothing to amend, nothing to force-push.
3. T2 · Turn end The Stop hook catches the agent declaring "done" with uncommitted code changes. It then runs make pre-commit before the turn can close.
4. T3 · Commit The git hook runs make pre-commit. That target calls detect-scope.sh to size the run and make charter-gates to run the diff sensors.

# The git hook is the ONLY gate on the commit path (commits.md) — the agent no
# longer runs `make pre-commit` explicitly. So a missing hook is not a small
# inconvenience, it is silent absence of verification. Warn loudly.
hooks_path="$(git config core.hooksPath || true)"
[ -z "$hooks_path" ] && hooks_path="$(git rev-parse --git-common-dir)/hooks"
if [ ! -x "${hooks_path}/pre-commit" ]; then
  echo "charter: ⚠ NO pre-commit hook at ${hooks_path}/pre-commit —
        commits are UNVERIFIED. Run \`make install-hooks\`."
fi

# Only intercept actual `git commit -m ...` invocations. Other commands
# (gh pr create, jira comment) may contain the literal string "git commit"
# in a body/heredoc — those must pass.
if ! echo "$cmd" | grep -qE '(^|[;&|[:space:]])git[[:space:]]+commit[[:space:]]'; then
  exit 0
fi

# Couldn't parse the message → fail open so unusual invocations pass.
[ -z "$msg" ] && exit 0

re='^(\[[A-Z]+-[0-9]+\] )?(feat|fix|refactor|test|docs|chore|perf|build|ci|style|revert)(\([a-z0-9._-]+\))?!?: .+$'

if [[ ! $subject =~ $re ]]; then
  charter: blocked git commit — subject does not match Conventional Commits.
    got: $subject
  Expected:  type(scope): subject
             [PROJ-123] type(scope): subject
  exit 2
fi

case "$f" in
    *.md|*.txt|*.rst|LICENSE|LICENSE.*) ;;
    # Generated by `make charter-index` and read by no checker, so `all`
    # would verify nothing a docs scope doesn't. Shell scripts deliberately
    # do NOT belong here: `make charter-gates` executes scripts/charter/*.sh
    # on every pre-commit, so a broken gate script silently stops gating.
    .claude/INDEX.json) ;;
    *.php)                       has_php=true   ;;
    *.ts|*.tsx|*.js|*.jsx|*.css) has_js=true    ;;
    *)                           has_other=true ;;
esac

The rule's why: and its "recent regression context" section do work no regex can. They tell the agent which kind of caution applies here. The caution is not "be careful". Pin behaviour before touching it, and never bundle a behaviour change with a refactor. Substitute your own module and your own incident; the shape holds.

---
description: The billing calculator is mission-critical and recently regressed
 : changes require regression tests, explicit approval for behavior changes,
  and characterization tests before any refactor
paths:
  - "app/Services/Billing/**"
  - "tests/**/Billing/**"
layer: project
locked: true
why: Billing regressions ship wrong money to the wrong accounts; the module
  already broke once in production this year and changes here have
  non-recoverable blast radius.
---
## IRON LAWS

**EVERY CHANGE PAIRED WITH A REGRESSION TEST.**
No change under `app/Services/Billing/` lands without a test under
`tests/UseCase/Services/Billing/` that fails *before* the change and passes
after. Pure refactors need characterization tests that pin current behavior.

**BEHAVIOR CHANGES REQUIRE EXPLICIT USER APPROVAL.**
Any diff that intentionally changes a calculator's output for a given input is
a behavior change — even if the new output is "more correct".
Do not bundle a behavior change with a refactor; separate commits, separate PRs.

## Recent regression context
The last production regression here turned on a subtle date/rounding
combination the suite did not pin. Until the characterization cards land,
assume every calculator branch is undertested.

The pairing, in both directions

1. Mechanical check-regression-test.sh: app code changed, no test file added or modified → exit 2. It cannot tell a good test from a bad one. It does not need to; it catches the case where there is none.
2. Judgment billing-risk.md: a read-only agent with allowed_tools restricted to Read, Grep, Glob and three git verbs. It reads the diff and asks whether the test actually pins the behaviour, and whether the math precision quietly shifted.
3. Mechanical check-diff-size.sh: warns past 15 files, blocks past 40. Pure counting; no idea what the card asked for.
4. Judgment scope-guard.md: reads the Jira card and the diff together. It flags drive-by edits that are individually reasonable and collectively out of scope.

# Blast-radius gate. A "small fix" that quietly rewrote 40 files is a signal to
# stop and look. Mechanical backstop to the intent-aware
# .claude/agents/scope-guard.md.

warn="${CHARTER_DIFF_WARN:-15}"
max="${CHARTER_DIFF_MAX:-40}"

if [ "$count" -gt "$max" ]; then
charter: blast-radius gate — diff touches $count files (hard ceiling $max).
A bounded change rarely rewrites this many files. Stop and look: is this
scope creep? Split it into separate cards, or — if the size is intentional
(a sweep, a rename) — re-run with CHARTER_BYPASS=1.
See .claude/agents/scope-guard.md for the intent-aware companion check.
fi

if [ "$count" -gt "$warn" ]; then
  echo "charter: blast-radius warning — diff touches $count files
        (soft limit $warn). Confirm it is all in scope for the card." >&2
fi

One card, start to finish

1. Drive /drive reads Jira, GitHub and git, never conversation memory, and derives the next phase from a table. Its iron law is DERIVE, DON'T REMEMBER.
2. Dispatch One subagent, foreground, one at a time. single-writer.md prohibits concurrent writers outright: no gate unlocks it, no urgency justifies it. The child gets identifiers (Jira key, PR number, worktree path), never a briefing. If dispatching a phase requires a briefing, the loop has drifted into holding state in conversation and must stop.
3. Fan out Inside /implement, four review agents run in parallel. That parallelism is legal because they are read-only sensors returning structured findings against a shared schema. Reconciliation is dedup-and-filter, not a forensic merge, and only the main loop applies fixes.
4. Smoke Exercise the feature's happy path against a running stack. Capture the evidence: a screenshot per meaningful state, or a report. Then re-read Jira, GitHub and git. Never trust the subagent's self-report: a reported success whose write is invisible in external state did not happen. And a change nobody drove is a change nobody proved. Both count as failed passes.
5. Price make drive-metrics KEY=PROJ-482: merged LOC against token cost, priced from a rate table. It reports without editorialising, and states its two limits. The arithmetic is list price, not an invoice, and LOC flatters a card that wrote a lot.
6. Learn /learn, invoked with no argument, runs the retrospective itself from the run's recorded artifacts. It proposes one rule diff, then stops for human approval.

Derive candidates from what the run *recorded*, never from recall:
1. A two-strike gate — the same failure twice.
2. A repeated `auto-resolved:` judgment call across checkpoint lines.
3. A review finding re-raised after a fix.
4. A workflow gap.
5. Cost — but expensive is not the same as wasteful, and only waste is a lesson.

Then apply the bar, and expect to fail it:
A candidate earns a charter change only if it recurred within the run,
or names a rule that was already wrong. A one-off call that went fine is
not a lesson — promoting it is exactly how a charter accumulates the
decorative rules `/prune` later has to retire.

"No lessons — nothing recurred" is a correct and common outcome.
Say it in one line and stop. Do not proceed to Step 1 to have something
to show for the invocation.

The brake

A flywheel that only adds rules is a ratchet, and every file the agent touches pays the cost. So the charter measures itself. make charter-index builds a descriptor index. Then make charter-which replays glob matching for any path. It totals the bytes of charter the agent reads before it reads the file. Run against one calculator file, on a charter of the size given at the top of this page:

Read the shape rather than the totals. Three rules load on **, so every file in the repo pays for them. Five more load on app/**/*.php, which is most of the backend. Only two are specific to this module. They are also the two smallest entries in the table. That distribution is the finding: rules that say nothing about this file consume almost the whole budget.

OVER load budget: 75830 / 60000 bytes across 15 charter files
  → merge into the rule that already owns this glob, or move reference
    material to .claude/features/ with narrower paths:
    (.claude/rules/charter-versioning.md)

Two read-only scans complete the brake. rule-drift.sh checks every paths: glob against the repo and flags globs matching zero files: a rule describing code that no longer exists. decorative-rule-check.sh counts each rule three ways: commits in 90 days, cross-references from skills and agents, mentions in CLAUDE.md/CHARTER.md. It marks the ones scoring zero everywhere with a headstone. Both scans run inside /prune, the deliberate counterpart to /learn.