@journeykit/canon 0.6.0 → 0.8.0

package/dist/skills/scripts/canon-retrieve-pre-dispatch.sh

@@ -0,0 +1,126 @@
+#!/bin/sh
+# canon-retrieve-pre-dispatch.sh — S8's generic pre-dispatch hook
+# (`s8-retrieve-before-task`, design.md decision 4, tasks.md task 3.1):
+# fires at a PreToolUse-equivalent point, generalized to any
+# `task`-shaped dispatch (Claude Code's `Task` tool / Codex's
+# equivalent subagent-dispatch tool call), mirroring a donor
+# harness's `pre-edit-pattern-lookup.ts` fail-soft, advisory-only
+# PreToolUse contract (design.md Context) — applied here to
+# role-scoped strategy retrieval instead of edit-pattern lookup.
+#
+# WIRING (design decision 4 — reuses S5's install-hooks mechanism
+# verbatim, invents no second convention): install the standard
+# `{matcher, hooks: [{type: "command", command, timeout}]}` entry via
+# the ALREADY-GENERIC `canon gate install-hooks` (S5, `s5-trust-spine-
+# gate`) — this script adds NO new installer code:
+#
+#   canon gate install-hooks --repo . --event PreToolUse \
+#     --matcher Task --command "sh /path/to/pre-dispatch.sh" \
+#     --timeout 15
+#
+# Place this file wherever your repo keeps materialized hook scripts
+# (e.g. `.claude/hooks/canon-retrieve-pre-dispatch.sh`, following the
+# manual-placement convention for materialized hook scripts)
+# and point `--command` at that path — same manual-
+# placement discipline canon-gate's own `PRE_COMMIT_SCRIPT` documents
+# ("install it directly as `.git/hooks/pre-commit`, or invoke it from
+# a lefthook/husky `run:` line").
+#
+# FAIL-SOFT, ADVISORY-ONLY — this script NEVER blocks the dispatch and
+# NEVER emits a `permissionDecision`; it only ever adds
+# `additionalContext` (Claude Code's `PreToolUse` hook output
+# contract) or stays silent. Every one of the following degrades to a
+# SILENT no-op, never a nonzero exit:
+#   - `canon` (or `jq`) not on PATH.
+#   - the dispatched tool is not task-shaped (`tool_name != "Task"`) —
+#     re-checked here even though the install recipe above already
+#     scopes the hook via `--matcher Task`, so this script stays a
+#     safe no-op if installed with a broader/no matcher.
+#   - no role derivable from the tool input, or empty/malformed stdin.
+#   - `canon retrieve` returns empty guidance — it is ALREADY
+#     fail-soft (S8Core's `retrieve_guidance` never errors; canon-
+#     cli's own CLI boundary never exits nonzero for a store outage,
+#     `crates/canon-cli/src/retrieve.rs`'s own module doc) — this
+#     script never adds a SECOND failure mode on top of that.
+#
+# ROLE/REGIME DERIVATION (an explicit design.md Open Question this
+# script resolves, pinned here): Claude Code's `Task` tool call
+# carries a `subagent_type` field naming which specialist role is
+# about to run — read directly as `--role` (kebab-cased). `--regime`'s
+# `<role>/<repo>/<area>/<hash>` is NOT fully recoverable from a bare
+# tool-call payload (S6's `area`/`hash` taxonomy is repo-specific, not
+# something a generic script can infer) — this ships a CONSERVATIVE
+# default (`area=general`, a stable hash of that area) a repo is
+# expected to override once it adopts a real regime taxonomy:
+#   - `CANON_RETRIEVE_ROLE`    overrides the auto-derived role.
+#   - `CANON_RETRIEVE_AREA`    overrides the default `general` area.
+#   - `CANON_RETRIEVE_HASH`    overrides the derived hash segment.
+#   - `CANON_RETRIEVE_REGIME`  overrides the whole assembled
+#                              `regime_key`, bypassing all of the above.
+set -eu
+
+command -v canon >/dev/null 2>&1 || exit 0
+command -v jq >/dev/null 2>&1 || exit 0
+
+INPUT="$(cat 2>/dev/null || true)"
+[ -n "$INPUT" ] || exit 0
+
+TOOL_NAME="$(printf '%s' "$INPUT" | jq -r '.tool_name // empty' 2>/dev/null || true)"
+[ "$TOOL_NAME" = "Task" ] || exit 0
+
+# `slugify` is used ONLY for the ROLE segment: `RoleId`'s grammar is
+# strict kebab (`[a-z0-9]+(-[a-z0-9]+)*`), and a kebab string is a
+# fixed point of `regime_key`'s own segment canonicalizer, so the
+# derived `--role` still equals the assembled regime_key's leading
+# segment. The repo/area segments are NOT slugified here — they are
+# canonicalized by `canon regime-key` below (the write-path serializer).
+slugify() {
+  printf '%s' "$1" | tr '[:upper:]' '[:lower:]' | tr -c 'a-z0-9' '-' | sed -E 's/-+/-/g; s/^-//; s/-$//'
+}
+
+ROLE="${CANON_RETRIEVE_ROLE:-}"
+if [ -z "$ROLE" ]; then
+  ROLE="$(printf '%s' "$INPUT" | jq -r '.tool_input.subagent_type // empty' 2>/dev/null || true)"
+fi
+ROLE="$(slugify "$ROLE")"
+[ -n "$ROLE" ] || exit 0
+
+REGIME="${CANON_RETRIEVE_REGIME:-}"
+if [ -z "$REGIME" ]; then
+  REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
+  REPO_RAW="$(basename "$REPO_ROOT")"
+  AREA_RAW="${CANON_RETRIEVE_AREA:-general}"
+  HASH="${CANON_RETRIEVE_HASH:-}"
+  if [ -z "$HASH" ]; then
+    HASH="$(printf '%s' "$AREA_RAW" | sha256sum 2>/dev/null | cut -c1-12)"
+    [ -n "$HASH" ] || HASH="$(printf '%s' "$AREA_RAW" | shasum -a 256 2>/dev/null | cut -c1-12)"
+  fi
+  [ -n "$HASH" ] || exit 0
+  # Assemble the regime_key through canon's OWN serializer (`canon
+  # regime-key` -> `canon_model::ids::regime_key`), the identical
+  # normalizer the S4/S6/S14 WRITE path uses — never a second shell
+  # derivation. A prior local `slugify` of the repo segment mapped a
+  # written `my_repo` to a queried `my-repo`, so `retrieve_guidance`
+  # fail-softed to empty and silently missed the namespace. `--role`
+  # is the already-kebab ROLE (a fixed point of that canonicalizer, so
+  # it still equals the assembled key's leading segment). Malformed
+  # segments make `canon regime-key` exit nonzero -> `|| true` -> empty
+  # REGIME -> the silent no-op just below.
+  REGIME="$(canon regime-key --role "$ROLE" --repo "$REPO_RAW" --area "$AREA_RAW" --hash "$HASH" 2>/dev/null || true)"
+  [ -n "$REGIME" ] || exit 0
+fi
+
+GUIDANCE_JSON="$(canon retrieve --role "$ROLE" --regime "$REGIME" --json 2>/dev/null || true)"
+[ -n "$GUIDANCE_JSON" ] || exit 0
+
+COUNT="$(printf '%s' "$GUIDANCE_JSON" | jq 'length' 2>/dev/null || echo 0)"
+[ "$COUNT" -gt 0 ] 2>/dev/null || exit 0
+
+CONTEXT="$(printf '%s' "$GUIDANCE_JSON" | jq -r --arg role "$ROLE" '
+  (["[canon-retrieve] guidance for role " + $role + ":"]
+   + (map("- " + .title + ": " + .content))) | join("\n")
+' 2>/dev/null || true)"
+[ -n "$CONTEXT" ] || exit 0
+
+jq -n --arg ctx "$CONTEXT" '{hookSpecificOutput: {hookEventName: "PreToolUse", additionalContext: $ctx}}'
+exit 0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@journeykit/canon",
-  "version": "0.6.0",
+  "version": "0.8.0",
   "description": "canon — harness knowledge substrate. JS launcher: platform detection + native binary resolution.",
   "license": "MIT",
   "type": "module",
@@ -22,14 +22,16 @@
     "directory": "packages/cli"
   },
   "scripts": {
-    "build": "bun build src/index.ts --outdir dist --target node --format esm",
+    "bundle-skills": "bun run scripts/bundle-skills.ts",
+    "build": "bun run bundle-skills && bun build src/index.ts --outdir dist --target node --format esm",
     "dev": "bun src/index.ts",
     "start": "node dist/index.js"
   },
   "dependencies": {},
   "optionalDependencies": {
-    "@journeykit/canon-core-darwin-arm64": "0.6.0",
-    "@journeykit/canon-core-linux-x64": "0.6.0"
+    "@journeykit/canon-core-darwin-arm64": "0.8.0",
+    "@journeykit/canon-core-linux-x64": "0.8.0",
+    "@journeykit/canon-core-win32-x64": "0.8.0"
   },
   "devDependencies": {
     "@types/node": "^20.0.0",