feat(document-to-markdown): read the formats Read cannot open, locally

Read handles PDF, images and notebooks natively; it does not open .docx,
.xlsx, .pptx, .epub, .msg or .zip. One uvx command does, with no install step
and no service.

Deliberately does NOT call docmark, which is the counter-intuitive part.
Measured 2026-09-24 on the same .docx, tables included: local markitdown and
the docmark service return byte-identical Markdown. Conversion of a
text-bearing document involves no LLM at all — the client is used only to
describe images. A network hop, a bearer token and an ingress for the same
bytes is not a trade worth making, and docmark is unreachable from a session
on koala anyway (ClusterIP, NetworkPolicy admits supervisor and dispatch only).

The boundary is measured, not asserted. The light extras register 18
converters; docmark registers 22. The difference is exactly DocxWithOCR,
PdfWithOCR, PptxWithOCR and XlsxWithOCR — the vision path, which is the one
case that must go through docmark because its sovereignty gate is the estate's
only audited route from an image to a model.

Two hazards the skill names because both bite silently:

  - markitdown[all] pulls az-doc-intel and az-content-understanding, Azure
    cloud converters. The extras list avoids having them on the path.
  - A blank template and a scanned PDF with no text layer both convert to zero
    characters and exit 0. Check the byte count; an empty success is the
    failure mode, and a confident summary of an unread document is the worst
    output this skill could produce.

Refs mathias/docmark#3

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "skills",
   "description": "Mathias's engineering skills library, plus the hooks and MCP servers that make it enforceable: TDD, clean code, debugging, specs, discovery, CI, and secret hygiene.",
-  "version": "0.19.1",
+  "version": "0.19.6",
   "author": {
     "name": "Mathias",
     "url": "https://git.d-ma.be/mathias"

.gitea/workflows/ci.yml

@@ -133,6 +133,33 @@ jobs:
             git add "$manifest"
             # The trailer below stops this push from re-triggering the tagger.
             git commit -m "chore(release): plugin.json version $version" -m "Skip-Release: true"
+
+            # skills#46. This commit is pushed with secrets.GITEA_TOKEN, and a
+            # push authenticated with the Actions token does NOT trigger
+            # workflows -- deliberate platform behaviour to stop CI recursing.
+            # `Skip-Release: true` stops the tagger recursing; nothing stopped
+            # the whole run, `gate` included, from never firing on this commit.
+            #
+            # So every release left main's HEAD with zero CI runs. Measured:
+            # 6cc224f (0.19.4) and f031c2b (0.19.5) each have none, and the
+            # koala cluster's CIHeadUncovered{repo=skills} had been firing on it
+            # for 85 hours -- an alert that can only be cleared by an unrelated
+            # human push.
+            #
+            # The commit is not cosmetic: plugin.json's version is what Claude
+            # Code gates plugin updates on, `task check` asserts it is not
+            # behind the latest tag, and the sed-and-retry logic above already
+            # shipped a mismatched manifest once (v0.10.0 reading 0.9.0).
+            #
+            # Gating it HERE rather than re-triggering CI afterwards: the commit
+            # is produced here, so it can be validated here, at the cost of one
+            # `task check` instead of a second full run -- and with no recursion
+            # risk to reintroduce.
+            if ! task check; then
+              echo "::error::the release-bump commit fails task check -- not pushing it" >&2
+              exit 1
+            fi
+
             if git push origin HEAD:main; then
               echo "bumped $manifest to $version"
               exit 0

SKILLS_INDEX.md

@@ -17,6 +17,7 @@ the SKILL.md, not the table.
 | `cognitive-load` | Cognitive complexity analysis (CLI score) | Identifying hotspots, before refactoring | claude-code, library |
 | `debug` | Hypothesis-first debugging discipline | Any test failure or bug whose cause is not immediately obvious — or a bug report with no failure text, which this skill refuses rather than guesses at | claude-code, library |
 | `discovery-framing` | Problem-before-solution discovery/framing method | Framing/validating a problem before any solution; opportunity sizing | claude-code, library |
+| `document-to-markdown` | Read .docx/.xlsx/.pptx/.epub/.msg and other formats the Read tool cannot open | You need the contents of a binary document — Word, Excel, PowerPoint, EPUB, Outlook .msg, a zip of documents — and Read returns nothing useful | claude-code, library |
 | `dream` | Reflective consolidation pass over an agent's memory directory | Pruning/merging memory files, rebuilding a memory index | claude-code, library, chat |
 | `experiment-spec` | Rigorous experiment spec — falsifiable hypothesis, quantitative acceptance, null-result protocol | Scoping a research/ML phase instead of a feature | claude-code, library |
 | `feature-spec` | Implementation-level feature spec (tighter than spec-driven-dev) | Scoping one feature inside an already-specced project | claude-code, library |

bin/check-mcp-env

@@ -41,7 +41,17 @@ if [ ! -f "$CONFIG" ]; then
 fi
 
 if [ ! -f "$ENVFILE" ]; then
+  # EXPECTED IN CI, and permanently so. The env-file holds live credentials, so
+  # it is not on the runner and must not be -- which means this check can only
+  # ever verify on a host that has one. infra#368 added jq to the runner image
+  # believing that would let CI verify; it removed the jq branch below and
+  # changed nothing else, because this branch is the binding one.
+  #
+  # Said plainly so nobody concludes CI covers this. It does not, it cannot,
+  # and putting ~/.op-env on the runner to make it would be trading a real
+  # secret for a green tick.
   echo "ERROR: no such env-file: $ENVFILE -- cannot verify MCP variables" >&2
+  echo "       (expected in CI: this check only verifies on a host with ~/.op-env)" >&2
   exit 2
 fi
 

bin/doctor

@@ -327,6 +327,44 @@ if [ "$ONLINE" -eq 1 ]; then
   fi
 fi
 
+# --- toolchain: golangci-lint must be built with Go >= GOROOT ---------------
+# infra#375, second recurrence. golangci-lint embeds the go/types export-data
+# reader from the Go it was BUILT with. When GOROOT is newer, that reader
+# cannot decode the stdlib and the linter dies:
+#
+#   panic: file requires newer Go version go1.27 (application built with go1.26)
+#
+# An earlier form printed a typecheck error naming the user's own file, which
+# sends people to debug code that is fine. It cost time in two repos before
+# anyone recognised it as a host fault rather than a repo fault.
+#
+# It is invisible to CI -- the runner image has matching versions -- so it
+# breaks exactly the local `task check` that AGENT.md requires before a push,
+# and nothing else. That is why it belongs in doctor: the signal was missing,
+# not the fix.
+#
+# Fails open on anything unparseable. A host with no Go toolchain is a normal
+# host, not a broken one, and doctor runs in a SessionStart hook everywhere.
+ver_lt() {
+  # 0 when $1 < $2 on major.minor. Deliberately not sort -V: this is POSIX sh
+  # that also runs on macOS, and the export-data format is minor-versioned, so
+  # patch level does not enter into it.
+  _a=${1%%.*}; _ar=${1#*.}; _am=${_ar%%.*}
+  _b=${2%%.*}; _br=${2#*.}; _bm=${_br%%.*}
+  case "$_a$_am$_b$_bm" in *[!0-9]*) return 1 ;; esac
+  [ "$_a" -lt "$_b" ] && return 0
+  [ "$_a" -gt "$_b" ] && return 1
+  [ "$_am" -lt "$_bm" ]
+}
+
+if command -v golangci-lint >/dev/null 2>&1 && command -v go >/dev/null 2>&1; then
+  lint_go=$(golangci-lint --version 2>/dev/null | sed -n 's/.*built with go\([0-9][0-9.]*\).*/\1/p' | head -1)
+  root_go=$(go version 2>/dev/null | sed -n 's/.*go version go\([0-9][0-9.]*\).*/\1/p' | head -1)
+  if [ -n "$lint_go" ] && [ -n "$root_go" ] && ver_lt "$lint_go" "$root_go"; then
+    add FAIL toolchain "golangci-lint built with go$lint_go but GOROOT is go$root_go — lint panics on every Go repo; reinstall it under the current toolchain (infra#375)"
+  fi
+fi
+
 # --- output -----------------------------------------------------------------
 if [ "$JSON" -eq 1 ]; then
   printf '{"checkout":"%s","expected":%s,"plugin":{"installed":"%s","available":"%s"},"findings":[%s]}\n' \

skills/document-to-markdown/SKILL.md

@@ -0,0 +1,126 @@
+---
+name: document-to-markdown
+surfaces: [claude-code, library]
+summary: "Read .docx/.xlsx/.pptx/.epub/.msg and other formats the Read tool cannot open"
+use_when: "You need the contents of a binary document — Word, Excel, PowerPoint, EPUB, Outlook .msg, a zip of documents — and Read returns nothing useful"
+description: Convert a binary document to Markdown locally with markitdown. No network, no LLM, safe for confidential material. Says when to use the docmark service instead.
+---
+
+# Document → Markdown
+
+## Overview
+
+`Read` opens PDFs, images and notebooks natively. It does **not** open `.docx`,
+`.xlsx`, `.pptx`, `.epub`, `.msg` or `.zip`. This skill converts those locally.
+
+One command, no install step, no service:
+
+```bash
+uvx --from 'markitdown[docx,pdf,pptx,xls,xlsx,outlook]' markitdown FILE > /tmp/out.md
+```
+
+First run downloads 38 packages (~10s); afterwards it is cached and instant.
+
+## When to use this
+
+You have a file on disk whose contents you need. That is the whole trigger.
+
+## When NOT to use this — use `docmark` instead
+
+Three cases, and only three:
+
+1. **A scanned PDF or an image whose content is pixels, not text.** Extracting
+   that needs a vision model. The estate's only audited path from an image to a
+   model is `docmark`, whose `sovereignty.py` constructs an LLM client *only*
+   when `DOCMARK_VISION_MODEL` matches `koala/*` or `iguana/*`, and silently
+   degrades to metadata-only otherwise. `tests/test_sovereignty.py` is the
+   load-bearing guard. Do not reach for a cloud vision model directly; that is
+   the boundary the gate exists to hold.
+2. **You are writing an in-cluster service** (something in `dispatch`,
+   `ingestion`, `cobalt-dingo`). Call `docmark` over MCP rather than bundling a
+   converter into each service. Its NetworkPolicy already admits `supervisor`
+   and `dispatch`.
+3. **The file is larger than ~25 MiB** or you want the input/output caps
+   enforced centrally.
+
+For a text-bearing document, the service returns **byte-identical output** to
+the local command — measured 2026-09-24 on the same `.docx`, tables included.
+A network hop, a bearer token and an ingress for the same bytes is not a
+trade worth making.
+
+## Confidential material
+
+**Conversion is local and involves no LLM.** `markitdown` uses an LLM client
+only to describe images; with none configured it does not call anything. No
+bytes leave the machine.
+
+That makes this safe for client material, unlike anything that routes through a
+model. Two cautions:
+
+- **Do not install `markitdown[all]`.** It pulls `az-doc-intel` and
+  `az-content-understanding` — Azure cloud document-intelligence converters.
+  They need explicit endpoint configuration to activate, but the extras list
+  above avoids having them on the path at all.
+- The output is a file like any other. If the source was confidential, so is
+  `/tmp/out.md` — and the usual rule applies: it must not reach an open-egress
+  sink (ADR-0014).
+
+## Verify the conversion, do not assume it
+
+The failure mode is **a successful exit with empty output**. A `.docx` template
+with no body content converts to zero characters and exits 0; so does a scanned
+PDF with no text layer. Both look like success.
+
+```bash
+uvx --from 'markitdown[docx,pdf,pptx,xls,xlsx,outlook]' markitdown FILE > /tmp/out.md
+wc -c /tmp/out.md
+```
+
+If the byte count is 0 or implausibly small for the document's size, the
+converter found no text layer — that is case 1 above, and it needs `docmark`
+with a vision model, not a retry.
+
+## What it converts
+
+18 converters with the extras above:
+
+```
+Audio, BingSerp, Csv, Docx, Epub, Html, Image, Ipynb, OutlookMsg, Pdf,
+PlainText, Pptx, Rss, Wikipedia, Xls, Xlsx, YouTube, Zip
+```
+
+`docmark` registers 22 — the same set plus `DocxWithOCR`, `PdfWithOCR`,
+`PptxWithOCR`, `XlsxWithOCR`. **Those four are the entire difference**, and
+they are exactly the vision path from case 1. Everything else is identical code
+running in a different place.
+
+## Reading the output
+
+Convert to a file and `Read` it, rather than piping into the transcript —
+a 200-page document produces a lot of Markdown and most of it will not be
+what you needed. Convert, then grep or read the section you want.
+
+## Provenance
+
+`docmark` ADR-0013 (`mathias/infra docs/decisions/`) owns the service; this
+skill is the local path. The split was decided in `mathias/docmark#3` on
+measurement, not preference: three of docmark's four plausible consumers are
+better served without it, and the fourth — vision under a sovereignty gate — is
+why it exists.
+
+## Escalation
+
+Local, read-only, single command. Nothing here is handed back mid-task.
+
+Stop and hand back when:
+
+- **The output is empty or near-empty.** Do not retry, and do not guess at the
+  contents. Say which file, what the byte count was, and that it needs a vision
+  path — then either use `docmark` with a local vision model, or ask. A
+  confident summary of a document you could not read is the worst possible
+  output of this skill.
+- **The document is client material and the task would send its contents
+  anywhere** — a commit message, an issue, a PR body, a cloud model. That is
+  ADR-0014's refusal, and it is not this skill's call to make.
+- **The file is encrypted or password-protected.** markitdown will fail; asking
+  for the password is a human's decision, not a retry.

tests/ci_gate_test.sh

@@ -53,5 +53,31 @@ printf '%s' "$tag_job" | grep -qE '^ +needs: *gate' \
   && ok 'ungated release.yml removed' \
   || no 'ungated release.yml removed'
 
+# skills#46. The manifest-bump commit is created and pushed by the `tag` job
+# using secrets.GITEA_TOKEN, and a push authenticated with the Actions token
+# does not trigger workflows -- deliberate platform behaviour to stop CI
+# recursing. So the ONE commit in this repo that no CI run has ever seen is the
+# one the release job writes, every release:
+#
+#   6cc224f  chore(release): plugin.json version 0.19.4   ->  0 runs
+#   f031c2b  chore(release): plugin.json version 0.19.5   ->  0 runs
+#
+# The koala cluster's CIHeadUncovered{repo=skills} alert had been firing on it
+# for 85 hours when this was found, and it cannot clear on its own -- only an
+# unrelated human push moves main's HEAD to a tested commit.
+#
+# The commit is not nothing: plugin.json's version is what Claude Code gates
+# plugin updates on, `task check` asserts it is not behind the latest tag, and
+# the sed-and-retry logic that writes it already shipped a mismatched manifest
+# once (v0.10.0 with the manifest reading 0.9.0).
+#
+# Anchored past comments: the tag job's prose already SAYS "`task check`
+# asserts the manifest is not behind the latest tag", and a bare grep matches
+# that and passes. A grep for a concept matches the rationale, and the more
+# carefully a thing is documented the more likely that is.
+printf '%s' "$tag_job" | grep -qE '^[^#]*task check' \
+  && ok 'the release-bump commit is gated before it is pushed' \
+  || no 'the release-bump commit is pushed with no gate -- nothing ever tests it (#46)'
+
 printf '\nci_gate: %d passed, %d failed\n' "$pass" "$fail"
 [ "$fail" -eq 0 ]

tests/doctor_test.sh

@@ -24,6 +24,18 @@ if [ ! -f "$DOCTOR" ]; then
   exit 1
 fi
 
+# The toolchain check (golangci-lint build-Go vs GOROOT) reads the real host
+# unless we say otherwise, which made every unrelated fixture fail on a machine
+# whose golangci-lint is genuinely stale. Give the whole suite a matching fake
+# pair so only the cases that OPT IN exercise that check.
+STUB_BIN="$WORK/_stub_bin"
+mkdir -p "$STUB_BIN"
+printf '#!/bin/sh\nprintf "golangci-lint has version 2.0.0 built with go1.27.0 from abc\\n"\n' > "$STUB_BIN/golangci-lint"
+printf '#!/bin/sh\ncase "$1" in version) printf "go version go1.27.0 linux/amd64\\n";; *) printf "\\n";; esac\n' > "$STUB_BIN/go"
+chmod +x "$STUB_BIN/golangci-lint" "$STUB_BIN/go"
+PATH="$STUB_BIN:$PATH"
+export PATH
+
 cleanup() { rm -rf "$WORK"; }
 trap cleanup EXIT INT TERM
 
@@ -291,5 +303,75 @@ else
   fail=$((fail + 1)); printf '  FAIL --online reports an unreachable remote as its own fault\n       rc=%s got: %s\n' "$_rc" "$_out"
 fi
 
+# --- toolchain: golangci-lint must be built with Go >= GOROOT ---------------
+# Measured on koala 2026-09-08 (infra#375, second recurrence): golangci-lint
+# 2.11.4 built with go1.26.1 against a go1.27.0 GOROOT PANICS on any Go repo --
+#
+#   panic: file requires newer Go version go1.27 (application built with go1.26)
+#
+# and the earlier form of the same fault printed a typecheck error naming YOUR
+# file, so the natural response is to debug code that is fine. It cost time in
+# two repos. CI never sees it (the runner image matches), so it breaks only the
+# local `task check` that AGENT.md requires before every push.
+#
+# The fixtures inject fake `go` and `golangci-lint` on PATH; nothing here reads
+# the real host toolchain, so the suite still passes on a machine without Go.
+mk_toolchain() { # $1=fixture  $2=lint-build-go  $3=goroot-go   ("" = absent)
+  _b="$1/bin"; mkdir -p "$_b"
+  if [ -n "$2" ]; then
+    printf '#!/bin/sh\nprintf "golangci-lint has version 2.11.4 built with %s from abc on 2026-01-01\\n"\n' "$2" > "$_b/golangci-lint"
+    chmod +x "$_b/golangci-lint"
+  fi
+  if [ -n "$3" ]; then
+    printf '#!/bin/sh\ncase "$1" in version) printf "go version %s linux/amd64\\n";; *) printf "\\n";; esac\n' "$3" > "$_b/go"
+    chmod +x "$_b/go"
+  fi
+  printf '%s' "$_b"
+}
+
+f=$(new_fixture golangci_old); wire_clean "$f"
+b=$(mk_toolchain "$f" go1.26.1 go1.27.0)
+_out=$(PATH="$b:$PATH" HOME="$f/home" DOCTOR_CHECKOUT="$f/ck" "$DOCTOR" 2>&1); _rc=$?
+if [ "$_rc" -ne 0 ] && printf '%s' "$_out" | grep -qi 'golangci'; then
+  pass=$((pass + 1)); printf '  ok   golangci-lint older than GOROOT is reported\n'
+else
+  fail=$((fail + 1)); printf '  FAIL golangci-lint older than GOROOT is reported\n       rc=%s got: %s\n' "$_rc" "$_out"
+fi
+
+# Matching versions are the normal case and must stay silent.
+f=$(new_fixture golangci_ok); wire_clean "$f"
+b=$(mk_toolchain "$f" go1.27.0 go1.27.0)
+_out=$(PATH="$b:$PATH" HOME="$f/home" DOCTOR_CHECKOUT="$f/ck" "$DOCTOR" 2>&1); _rc=$?
+if [ "$_rc" -eq 0 ]; then
+  pass=$((pass + 1)); printf '  ok   golangci-lint matching GOROOT is silent\n'
+else
+  fail=$((fail + 1)); printf '  FAIL golangci-lint matching GOROOT is silent\n       rc=%s got: %s\n' "$_rc" "$_out"
+fi
+
+# A NEWER lint than GOROOT is fine -- the constraint is one-directional.
+f=$(new_fixture golangci_newer); wire_clean "$f"
+b=$(mk_toolchain "$f" go1.28.0 go1.27.0)
+_out=$(PATH="$b:$PATH" HOME="$f/home" DOCTOR_CHECKOUT="$f/ck" "$DOCTOR" 2>&1); _rc=$?
+if [ "$_rc" -eq 0 ]; then
+  pass=$((pass + 1)); printf '  ok   golangci-lint newer than GOROOT is not a fault\n'
+else
+  fail=$((fail + 1)); printf '  FAIL golangci-lint newer than GOROOT is not a fault\n       rc=%s got: %s\n' "$_rc" "$_out"
+fi
+
+# Fail OPEN on anything unparseable. doctor runs in a SessionStart hook on every
+# machine; a toolchain it cannot read is not a fault it may assert. (A host with
+# no Go at all takes the same path -- command -v simply misses.)
+f=$(new_fixture golangci_garbled); wire_clean "$f"
+_b="$f/bin"; mkdir -p "$_b"
+printf '#!/bin/sh\nprintf "golangci-lint dev build\\n"\n' > "$_b/golangci-lint"
+printf '#!/bin/sh\nprintf "go version devel\\n"\n' > "$_b/go"
+chmod +x "$_b/golangci-lint" "$_b/go"
+_out=$(PATH="$_b:$PATH" HOME="$f/home" DOCTOR_CHECKOUT="$f/ck" "$DOCTOR" 2>&1); _rc=$?
+if [ "$_rc" -eq 0 ]; then
+  pass=$((pass + 1)); printf '  ok   an unreadable toolchain version is not asserted as a fault\n'
+else
+  fail=$((fail + 1)); printf '  FAIL an unreadable toolchain version is not asserted as a fault\n       rc=%s got: %s\n' "$_rc" "$_out"
+fi
+
 printf '\ndoctor: %d passed, %d failed\n' "$pass" "$fail"
 [ "$fail" -eq 0 ]

tests/pre_task_ritual_test.sh

@@ -0,0 +1,103 @@
+#!/bin/sh
+# Tests for hooks/pre-task-ritual.sh — the UserPromptSubmit hook that surfaces
+# rule 0 on every prompt (#44).
+#
+# WHY THIS EXISTS, AND WHAT IT DELIBERATELY DOES NOT CLAIM
+# This hook is three lines and echoes a string. It has no bash-isms, so it
+# cannot fail the way secret-guard appeared to; that risk was overstated when
+# this test was requested. What it CAN do is stop working silently, and its
+# whole job is to be visible -- if the text stops appearing, rule 0 stops being
+# surfaced on every prompt and nothing announces that.
+#
+# THE ASSERTION THAT MATTERS IS REGISTRATION.
+# A hook file that exists but is not in hooks.json does absolutely nothing, and
+# looks identical on disk to one that works. So this derives the command from
+# hooks.json rather than assuming one: the registration is the source of truth
+# for how the hook is invoked, and a test that invokes it its own way is
+# testing something the harness never runs.
+#
+# That distinction is not hypothetical. secret_guard_test.sh piped its hook into
+# `sh` while hooks.json runs it with `bash`; on Debian /bin/sh is dash, which
+# has no `set -o pipefail`, and the suite failed 20/20 in CI against a hook that
+# was fine in production. The test was wrong, not the hook. Deriving the
+# invocation removes the opportunity for that class of mistake entirely.
+
+set -u
+
+ROOT=$(CDPATH= cd -- "$(dirname -- "$0")/.." && pwd)
+HOOK="$ROOT/hooks/pre-task-ritual.sh"
+REG="$ROOT/hooks/hooks.json"
+
+pass=0
+fail=0
+ok() { pass=$((pass + 1)); printf '  ok   %s\n' "$1"; }
+no() { fail=$((fail + 1)); printf '  FAIL %s\n' "$1"; }
+
+if ! command -v jq >/dev/null 2>&1; then
+  printf '  SKIP: jq not installed -- cannot read the hook registration\n'
+  exit 0
+fi
+
+[ -f "$HOOK" ] || { printf '  FAIL %s does not exist\n' "$HOOK"; exit 1; }
+ok 'hook file exists'
+
+[ -f "$REG" ] || { printf '  FAIL %s does not exist\n' "$REG"; exit 1; }
+
+# 1. Registered, and on the right event. UserPromptSubmit is what makes it fire
+#    on every prompt; registered under any other event it would still "exist"
+#    and still do nothing useful.
+CMD=$(jq -r '
+  .hooks.UserPromptSubmit[]?.hooks[]?
+  | select(.command | test("pre-task-ritual"))
+  | .command
+' "$REG" 2>/dev/null | head -1)
+
+if [ -n "$CMD" ]; then
+  ok 'registered on UserPromptSubmit in hooks.json'
+else
+  no 'not registered on UserPromptSubmit -- the file would exist and never run'
+  printf '  %d passed, %d failed\n' "$pass" "$fail"
+  exit 1
+fi
+
+# 2. Run it EXACTLY as registered, with the harness variable substituted the way
+#    the harness substitutes it.
+OUT=$(CLAUDE_PLUGIN_ROOT="$ROOT" sh -c "$CMD" 2>&1)
+rc=$?
+
+[ "$rc" -eq 0 ] && ok 'exits 0 as registered' \
+  || no "exited $rc -- a non-zero UserPromptSubmit hook can block the prompt"
+
+# 3. The content is the point. A hook that runs and says nothing is the same as
+#    no hook, and that is the failure this pins.
+if [ -n "$OUT" ]; then
+  ok 'produces output'
+else
+  no 'produced no output -- rule 0 would silently stop being surfaced'
+fi
+
+case "$OUT" in
+  *"Rule 0"*) ok 'names rule 0' ;;
+  *)          no "does not name rule 0: '$OUT'" ;;
+esac
+
+# Each of the four steps, individually. Naming rule 0 while having quietly lost
+# a step is the plausible drift, not losing the whole line.
+for step in brain_query skill 'failing test' 'success criterion'; do
+  case "$OUT" in
+    *"$step"*) ok "states the '$step' step" ;;
+    *)         no "lost the '$step' step from the ritual" ;;
+  esac
+done
+
+# 4. The interpreter named in the registration must actually exist, or the hook
+#    fails on that host with a message nobody reads.
+INTERP=${CMD%% *}
+if command -v "$INTERP" >/dev/null 2>&1; then
+  ok "registered interpreter '$INTERP' is available"
+else
+  no "registered interpreter '$INTERP' not found on this host"
+fi
+
+printf '  %d passed, %d failed\n' "$pass" "$fail"
+[ "$fail" -eq 0 ]

tests/secret_guard_test.sh

@@ -24,7 +24,7 @@ decision_for() {
   _tool=${2:-Bash}
   jq -n --arg t "$_tool" --arg c "$_cmd" \
     '{tool_name: $t, tool_input: {command: $c}}' \
-    | sh "$GUARD" \
+    | "$GUARD" \
     | jq -r '.hookSpecificOutput.permissionDecision // "MISSING"'
 }
 
@@ -42,6 +42,20 @@ expect() {
   fi
 }
 
+# Invoked as a PROGRAM above, never as `sh "$GUARD"`. Claude Code executes the
+# hook directly, so its `#!/usr/bin/env bash` shebang decides the interpreter;
+# piping it into `sh` overrides that and tests something production never runs.
+#
+# It is not academic. On Debian /bin/sh is dash, which has no `set -o
+# pipefail`, so the guard died on line 23 before reading a byte of input and
+# every deny case returned empty -- 20 of 20 assertions failed in CI while
+# passing on a host where /bin/sh happens to be bash. A security gate that
+# tests green and cannot start is the worst of both.
+if [ ! -x "$GUARD" ]; then
+  printf 'FAIL: %s is not executable -- it must run via its own shebang\n' "$GUARD"
+  exit 1
+fi
+
 if [ ! -f "$GUARD" ]; then
   printf 'FAIL: %s does not exist\n' "$GUARD"
   exit 1