holocron/plugins/kyberforge/docs/hooks.md

Hooks

Reference for this plugin's hook definitions: where to edit them, what apm does with them, and what the host ends up reading.

Where to edit

Author hooks in plugins/kyberforge/.apm/hooks/*.json. .apm/ is the only content source and apm install is the only supported install path (ADR-0024) — there is no generated mirror at the plugin root and no per-plugin plugin.json, so .apm/hooks/ is both where you edit and what ships.

For Claude, apm merges the event bindings from every *.json in that directory into the consuming project's .claude/settings.json; see "Deployed shape" below. That merge is Claude's rendering, not a general rule: Copilot gets one file per source file (see "GitHub Copilot CLI and Codex"). Scripts a hook invokes live in the same directory, alongside the JSON that references them.

Hook file structure

The shape Claude Code reads, and therefore the shape to author under .apm/hooks/:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          { "type": "command", "command": "echo 'tool used'", "timeout": 10 }
        ]
      }
    ]
  }
}

Events: see plugins/kyberforge/docs/research/docs/microsoft-apm/hooks-primitive-schema.md, section "Events: _HOOK_EVENT_MAP", for which names apm renames per target and which it passes through unchanged. For Claude, author every event in PascalCase (SessionStart, UserPromptSubmit, PreCompact, and so on). A camelCase name apm does not map, such as userPromptSubmit, draws only a non-fatal warning and never fires; an all-lowercase one draws no warning at all and never fires either.

Referencing a script — use the .apm/ path

Use ${PLUGIN_ROOT} to reference scripts inside this plugin; the plugin runs from a cache or apm_modules/ path after install, not its original repo location. ${PLUGIN_ROOT} is apm's target-neutral token; apm rewrites it exactly as it rewrites ${CLAUDE_PLUGIN_ROOT} — verified byte-identical in the deployed .claude/settings.json with apm 0.28.0 — so prefer it, and factory-audit suggests it. Address the script at its .apm/ path:

"command": "${PLUGIN_ROOT}/.apm/hooks/check-apm-current.sh"

The obvious-looking ${PLUGIN_ROOT}/hooks/check-apm-current.sh does not work, and fails quietly enough to be worth spelling out. apm resolves the placeholder against the installed package root, and there is no hooks/ directory there at all — the package's content is .apm/. apm prints Hook script not found: .../hooks/check-apm-current.sh and then deploys the hook anyway, pointing at a path with no file behind it.

tests/test-apm-current-hook.sh pins the reference so a well-meaning "simplification" back to hooks/ fails the suite rather than silently disabling the hook.

Deployed shape

At install, apm merges the event bindings into .claude/settings.json, copies the referenced script to .claude/hooks/<pkg>/ (preserving its executable bit, preserving the .apm/hooks/ subpath), and rewrites command to a ${CLAUDE_PROJECT_DIR}-relative path. Ownership of its own entries is tracked in a .claude/apm-hooks.json sidecar, so an uninstall removes them without touching hand-authored hooks. .claude/hooks/ is gitignored install output; the sidecar is committed alongside .claude/settings.json, because without it a fresh clone's apm install treats the committed entry as user-owned and adds a duplicate, and apm audit --ci reports drift (ADR-0019, correction 2026-09-16).

Note that apm's executable-trust gate is off unless the consuming project's apm.yml has an executables: block — without one, package hooks deploy with no prompt. The allow key carries a version (kyberforge#<version>), but apm 0.28.0 matches grants version-blind, so a version bump on one side does not stop the hook deploying. check-executables-allow-sync is a pre-push gate for this repo's own convention that the key tracks plugins/kyberforge/apm.yml's version:, not for an apm mechanic. See ADR-0019, correction 2026-09-19, and the comment above executables: in the root apm.yml.

The SessionStart hook

check-apm-current.sh keeps an apm-consumed install level with its remote: it runs apm outdated, and if anything is behind, runs apm update --yes and returns reloadSkills: true so the running session picks up the redeployed content. reloadSkills is a documented SessionStart hookSpecificOutput field that makes Claude Code re-scan skill directories once the hooks finish (code.claude.com/docs/en/hooks, checked 2026-09-29). Rationale, measurements, and the failure modes are in ADR-0019.

Claude Code only, and where it looks for the lockfile. The hook exits 0 at once, silently and without calling apm, unless CLAUDE_PROJECT_DIR is set and non-empty — Claude Code exports it for SessionStart hooks, and Copilot and Codex do not document setting it, so the guard keeps the hook inert there unless the variable is inherited from the user's environment (see below). It then takes ${CLAUDE_PROJECT_DIR} as the project directory and exits silently unless that directory holds an apm.lock.yaml — which is what makes it inert in any project that does not consume packages through apm. Both apm invocations run against the same directory. The earlier spelling checked a bare apm.lock.yaml against the session's cwd, so a session opened in a subdirectory of an apm-consuming repo no-opped silently. Do not reintroduce a cwd fallback: under a host that sets no CLAUDE_PROJECT_DIR the lockfile guard passes in every apm consumer, and the fallback ran apm update --yes there (ADR-0019, correction 2026-09-28).

The timeout in hooks.json must exceed the script's own budget. The script spends at most timeout -k 5 60 apm outdated plus timeout -k 5 300 apm update, 370 s counting each 5 s SIGKILL grace; the hook entry declares timeout: 380, the sum plus a buffer. Set it lower and a slow remote gets the hook SIGKILLed mid-apm update, leaving a partially redeployed .claude/skills/ and emitting no notice — precisely the silent failure the hook exists to prevent. tests/test-apm-current-hook.sh pins the relationship (host timeout > sum of the script's internal timeouts) rather than the literal, so raising either side alone fails the suite.

The time limits are portable and hard to escape (ADR-0019, amendment 2026-09-29). The script uses timeout, or gtimeout where only Homebrew coreutils provides it (stock macOS). With neither, it emits a notice and exits without running apm, instead of dying silently on exit 127. It exports GIT_TERMINAL_PROMPT=0, so a remote that wants credentials fails at once instead of waiting out the timeout on a prompt nobody can see. Where flock exists and apm_modules/ does too, apm update runs under a non-blocking lock on apm_modules/.kyberforge-apm-update.lock. A second session that starts during a refresh skips its own and says so. Without flock the refresh runs unserialised.

The lock advice is neutral when the default branch is unknown. The notice tells you to discard the rewritten apm.lock.yaml on a feature branch and to decide deliberately on the default branch. It learns the default from refs/remotes/origin/HEAD, which git remote add never writes. When that ref is unset, on a detached HEAD, or outside a git checkout, it says only "commit it or discard it deliberately" rather than guessing main (ADR-0019, amendment 2026-09-19).

Staleness is detected by matching apm's summary line, and both spellings count. apm outdated has no --json or otherwise machine-readable output (verified against apm 0.28.0), so the hook greps its text. apm prints 1 outdated dependency found in the singular when exactly one package is behind and N outdated dependencies found otherwise; matching only the plural silently misses a one-package drift. Because a mocked apm would keep a reworded release invisible, the test suite stages a genuinely outdated dependency against the real apm — a local git repo reached through url.<path>.insteadOf rewrites, so it needs no network — and replays that genuine output through the hook.

GitHub Copilot CLI and Codex

apm writes this plugin's hook for Copilot and Codex too; whether they run it is unverified. kyberforge's apm.yml declares targets: [claude, copilot, codex], and targets: is package-wide, so the hook reaches every target the package does (plugins/kyberforge/docs/research/docs/microsoft-apm/hooks-primitive-schema.md, verified against apm 0.28.0):

• Copilot gets .github/hooks/kyberforge-hooks.json, one file per source file. apm renames the event (SessionStart → sessionStart), rewrites the script path, adds version: 1, and otherwise passes the nested Claude shape through — it does not reshape it into Copilot's flat bash/powershell/timeoutSec form. Whether Copilot CLI executes a nested entry, or honours matcher, has not been verified.
• Codex gets the entry merged into .codex/hooks.json, but only when .codex/ already exists; otherwise nothing is written.

This is accepted rather than fixed (ADR-0019, amendment and correction 2026-09-28). The hook's behaviour is Claude-specific anyway — the startup matcher, CLAUDE_PROJECT_DIR, and the reloadSkills output — and a harness that does run it exits immediately, because the script's first guard exits 0 when CLAUDE_PROJECT_DIR is unset. The apm.lock.yaml guard cannot do that job: apm install wrote the lock, so it passes in every project the hook reaches. The only apm-native way to keep it Claude-only is a separate package whose apm.yml declares target: claude; per-file target routing (claude-hooks.json) is deprecated, and kyberforge cannot narrow its own targets: without dropping its skills from Copilot and Codex.

An earlier version of this section said Copilot loads no hooks from this plugin, because nothing could point Copilot at a hooks file and apm did no per-target shaping. Both halves are superseded: apm deploys the file into Copilot's hooks directory itself, and does rename events per target.

Symlinks under .apm/ do not survive, and nothing reports it

Do not author any file under plugins/kyberforge/.apm/ as a symlink. apm's copy path filters symlinks out silently: ignore_non_content() in apm_cli/security/gate.py is a shutil.copytree ignore callback that drops every entry answering is_symlink(), commented "Excludes symlinks (security)". The file never reaches the consumer's install, and no warning is printed at any point.

No gate catches this. The pre-push check that used to read the .apm/ tree and fail on the offending path was deleted along with the content mirror, and the decision was taken not to replace it (ADR-0024 consequence 7). This document is the only thing standing between a symlink and silent content loss. No symlink exists under any .apm/ today; add one and it is dropped on deploy with nothing reporting it. Replace it with a regular file.

The old exemption for .apm/<category>/<name>/tests/ no longer applies either. That directory was exempt only because the mirror excluded it; apm deploys it like any other content (ADR-0024 consequence 2), so a symlink there loses content the same as anywhere else.