Files
stack/packages/mosaic/framework/tools/git/wrapper-guard.sh
T
Hermes Agent e3a0ee87b3 framework: make tool discoverability, workspace placement and model tiering mechanical
An undocumented tool is, from inside an agent session, indistinguishable from a
tool that was never written. The framework shipped 26 git wrappers and named 6 of
them in its resident index docs — 23% discoverability, with pr-review.sh among the
missing. The observable consequence was an agent obeying Constitution gate 7 as
best it could see it, reaching for raw curl, sending GitHub's APPROVE to a Gitea
host, and getting HTTP 200 with the review silently filed PENDING. Three times.
That is not a discipline failure and no amount of prose fixes it.

Four changes, each converting a rule that decayed into a mechanism that cannot:

- check-tools-index.sh (new, CI-blocking): every tool in an enforced suite must be
  named in a resident index doc, and every tool an index names must exist. The git
  suite is enforced now; other suites report coverage without failing, so the
  ratchet tightens one reviewed PR at a time instead of landing as one sweep. The
  enforced list is framework-owned rather than a marker inside operator-owned
  TOOLS.md — a doc marker would let an operator silence the gate on exactly the
  host where it matters most. Carries --self-test, because a checker that only
  ever passes is indistinguishable from one that is not running.

- TOOLS-REFERENCE.md: complete 28-entry git index, plus the APPROVED/APPROVE
  dialect note that explains why pr-review.sh is not a formality.

- mosaic-worktree.sh + wrapper-guard.sh (upstreamed): the rule "big work goes on a
  work filesystem" already existed in prose, and 255 GB accumulated in $HOME across
  842 directories anyway, under five simultaneous placement conventions on one
  host. The helper therefore exposes no placement decision — given a branch name,
  every path is derived from `git worktree list --porcelain`. Worktrees rather than
  clones because enumerability is the only thing that makes reclaim safe, and
  reclaim is by evidence (clean tree + no unpushed commits), never by size or age.
  The guard blocks three mechanically-detectable mistakes and nothing else:
  a checkout into $HOME, a raw provider-API write to an endpoint that has a
  wrapper, and the literal APPROVE event. Reads pass untouched.

- STANDARDS.md: model tiering as a standard, named by capability class so it
  survives a model generation. Start cheapest, escalate on evidence, benchmark
  before demoting a task class, and keep the class->model binding in operator
  config with the DB-backed config service as the end state.

Registering the guard in runtime/claude/settings.json is the point of upstreaming
it: ~/.claude/settings.json is a framework-managed copy, so a hand-added hook there
is destroyed by the next upgrade. In the template it survives, and it reaches every
host instead of one.
2026-08-12 16:51:17 -05:00

136 lines
5.8 KiB
Bash
Executable File

#!/usr/bin/env bash
# wrapper-guard.sh — PreToolUse hook on Bash.
#
# Blocks three specific, mechanically-detectable mistakes that prose has
# repeatedly failed to prevent:
#
# 1. A checkout (git clone / git worktree add) targeting $HOME.
# Root cause of a fleet host's /home filling to 100% — 255 GB, 842 dirs.
#
# 2. A raw provider API WRITE against an endpoint that already has a Mosaic
# wrapper. Constitution gate 7 requires the wrapper; the wrapper knows
# provider dialect, identity, and queue-guard ordering that raw curl does
# not. Reads are untouched — they are how you gather evidence.
#
# 3. The literal review event "APPROVE". Gitea's vocabulary is APPROVED;
# it accepts APPROVE with HTTP 200, silently files the review PENDING,
# and then 422s on submit. This one is unconditionally wrong on Gitea and
# is what a verdict silently failing to land looks like.
#
# Design constraint: this hook must not become something agents route around.
# It blocks WRITES to endpoints with a known wrapper, and nothing else. Raw
# curl for reads, for registry/manifest calls, and for endpoints with no
# wrapper (there are many) all pass untouched.
#
# Break-glass, for a genuine gap where no wrapper can express the call:
# MOSAIC_WRAPPER_OVERRIDE=1 <command>
# Using it means "no wrapper covers this" — if that is wrong, the fix is to
# extend the wrapper, not to keep typing the override.
#
# Exit codes (Claude Code PreToolUse): 0 = allow, 2 = block with message.
set -euo pipefail
INPUT="$(cat)"
CMD="$(printf '%s' "$INPUT" | jq -r '.tool_input.command // empty' 2>/dev/null || true)"
[ -z "$CMD" ] && exit 0
# Honour the override only when it is set in the command itself or the env.
case "$CMD" in *MOSAIC_WRAPPER_OVERRIDE=1*) exit 0 ;; esac
[ "${MOSAIC_WRAPPER_OVERRIDE:-0}" = "1" ] && exit 0
W="$HOME/.config/mosaic/tools/git"
# ---- 1. checkout into $HOME ------------------------------------------------
if printf '%s' "$CMD" | grep -Eq 'git[^|;&]*(clone|worktree[[:space:]]+add)'; then
# Any argument that resolves under $HOME and is not under a work filesystem.
if printf '%s' "$CMD" | grep -Eq "(^|[[:space:]=\"'])(~|\\\$HOME|$HOME)/"; then
cat <<EOF
BLOCKED: this checks a repository out under \$HOME.
\$HOME holds configuration, credentials, state and caches. It does not hold
checkouts, worktrees, scratch files, or build output. One fleet host's /home hit
100% (394 G) with 255 GB of agent workspaces accumulated exactly this way.
Use the helper, which derives the path so you do not have to choose one:
~/.config/mosaic/tools/git/mosaic-worktree.sh new <branch> # /src/<repo>-worktrees/<slug>
~/.config/mosaic/tools/git/mosaic-worktree.sh path <branch> # show where it would go
~/.config/mosaic/tools/git/mosaic-worktree.sh rm <branch> # removal is part of the task
Worktrees, not clones: they share the object store, and \`git worktree list\`
makes every one of them enumerable — which is the only reason cleanup can
ever be safe.
EOF
exit 2
fi
fi
# ---- 2/3. provider API writes ---------------------------------------------
# Only consider calls that are (a) to a provider API path and (b) mutating.
is_api=0
printf '%s' "$CMD" | grep -Eq '/api/v1/repos/|api\.github\.com/repos/' && is_api=1
if [ "$is_api" -eq 1 ]; then
is_write=0
printf '%s' "$CMD" | grep -Eq -- '-X[[:space:]]*(POST|PATCH|PUT|DELETE)|--request[[:space:]]*(POST|PATCH|PUT|DELETE)' && is_write=1
# curl sends POST implicitly when given a body.
printf '%s' "$CMD" | grep -Eq -- '--data|-d[[:space:]]' && is_write=1
if [ "$is_write" -eq 1 ]; then
endpoint=""; wrapper=""
case "$CMD" in
*"/pulls/"*"/reviews"*|*"/pulls/"*"/requested_reviewers"*)
endpoint="pull-request review"; wrapper="$W/pr-review.sh" ;;
*"/pulls/"*"/merge"*) endpoint="pull-request merge"; wrapper="$W/pr-merge.sh" ;;
*"/issues/"*"/comments"*) endpoint="issue comment"; wrapper="$W/issue-comment.sh" ;;
*"/pulls"*) endpoint="pull request"; wrapper="$W/pr-create.sh" ;;
*"/issues"*) endpoint="issue"; wrapper="$W/issue-create.sh" ;;
*"/milestones"*) endpoint="milestone"; wrapper="$W/milestone-create.sh" ;;
esac
if [ -n "$wrapper" ] && [ -x "$wrapper" ]; then
cat <<EOF
BLOCKED: raw provider API write to the $endpoint endpoint.
A Mosaic wrapper already covers this and the Constitution (gate 7) requires it
before any raw provider call:
$wrapper
The wrappers are not a formality. They carry provider-dialect differences that
raw curl silently gets wrong — Gitea's review event is APPROVED, GitHub's is
APPROVE, and Gitea accepts the wrong one with HTTP 200 while filing the review
as PENDING. They also resolve identity explicitly, which matters on a host
where the default login is an admin account.
Run \`$(basename "$wrapper") --help\` for the flags.
If no wrapper flag can express this call, that is a wrapper gap: extend the
wrapper. To proceed anyway for a genuine gap, prefix MOSAIC_WRAPPER_OVERRIDE=1.
EOF
exit 2
fi
fi
fi
# ---- 3. the APPROVE/APPROVED trap, wherever it appears ---------------------
if printf '%s' "$CMD" | grep -Eq '"event"[[:space:]]*:[[:space:]]*"APPROVE"'; then
cat <<EOF
BLOCKED: review event "APPROVE" is not valid on Gitea.
Gitea's vocabulary is "APPROVED". It accepts "APPROVE" with HTTP 200, silently
files the review as PENDING, and then fails the submit endpoint with
422 "review stay pending" — so the verdict looks placed and is not.
("REQUEST_CHANGES" is spelled identically on both providers; only the approve
path carries this trap.)
Use $W/pr-review.sh, which sends the correct token for the detected provider.
Whatever you use, re-read GET /pulls/{n}/reviews and assert state==APPROVED
before reporting a verdict placed.
EOF
exit 2
fi
exit 0