@codyswann/lisa 4.54.13 → 4.55.0

package/plugins/lisa/hooks/guard-dedupe.bash

@@ -0,0 +1,337 @@
+# shellcheck shell=bash
+# Evaluate a dual-channel guard once per tool call, without de-registering it.
+#
+# Sourced — never executed — by every guard that Lisa registers on BOTH of its
+# enforcement channels: the repository dispatcher
+# (`scripts/lisa-enforcement-fallback.sh`, which sweeps the guard roster in one
+# process) and the plugin manifest, which registers the same guards
+# individually. On a machine where both are live the harness runs each guard
+# twice for one tool call. The verdicts agree — the guards are idempotent — so
+# the defect is cost, not correctness (CodySwannGT/lisa#3814).
+#
+# WHY THIS IS NOT "DROP ONE REGISTRATION".
+#
+# The obvious fix is to de-register the fallback where the plugin covers the
+# same ground. That fix cannot be made safely from here, and the reason is
+# written into the dispatcher itself: plugin-hook LIVENESS is not observable
+# from disk. `installed_plugins.json` answers "was this plugin ever installed",
+# which is project-blind, enablement-blind and session-blind — hooks load at
+# session start, the record is rewritten on any install. A stand-down keyed on
+# that record switched enforcement off in sessions where the plugin hooks were
+# not loaded, and did it silently. So the dispatcher runs unconditionally and
+# always will.
+#
+# What IS observable is EXECUTION. A guard that has already run in this session
+# from a different directory is proof — not inference — that a second channel is
+# live right now. This file turns that observation into a per-call memo, so the
+# second copy of an identical guard replays a verdict the first one produced
+# instead of computing it again. Nothing is de-registered, no matcher narrows,
+# and a host with the plugin absent takes the single-channel path where the memo
+# can never hit.
+#
+# WHAT THE MEMO IS ALLOWED TO SHORT-CIRCUIT, and the three conditions that make
+# it sound. Every one of them is a REFUSAL to dedupe when it cannot be proven:
+#
+#   1. BYTE-IDENTICAL CODE. The digest covers the guard script and this library
+#      verbatim. Two channels at different vintages therefore compute different
+#      keys and both run — which is required, because the agent is meant to see
+#      the UNION of two vintages' verdicts, and a tightening on either channel
+#      must take effect at once. Deduplicating across vintages would let the
+#      older, weaker copy answer for the newer one.
+#   2. IDENTICAL PAYLOAD, SAME TOOL CALL. The digest covers the payload and a
+#      per-call discriminator — the transcript's size, which grows with every
+#      turn. Without the discriminator a command repeated later in the session
+#      would hit a memo written before the repository changed underneath it, and
+#      a guard that classifies against the working tree would be answering with
+#      a stale verdict. No discriminator available means no memo.
+#   3. ALLOW VERDICTS ONLY. A refusal is never memoised and never replayed. That
+#      is deliberate: refusals are the rare path, so deduplicating them saves
+#      nothing measurable, and it removes the whole class of hazard the ticket
+#      names — a replayed refusal whose producing channel died mid-run, or a
+#      refusal reaching the agent without the message that explains it. On a
+#      refusal every channel evaluates and speaks exactly as it does today.
+#
+# COST ON THE PATH THAT GETS NO BENEFIT. A host with one channel must not pay
+# for a mechanism it cannot use. The channel check is builtins only — one small
+# file read, no subshell, no fork — and returns before the digest is computed
+# unless a second channel has actually been seen. The digest, and the one `stat`
+# that feeds it the per-call discriminator, are reached only where a second
+# channel is live, which is exactly where they are repaid.
+
+# Key this guard would memoise under, once computed. Empty means "record
+# nothing on exit".
+lisa_guard_memo_key=""
+
+# File that memo is written to.
+lisa_guard_memo_file=""
+
+# Base directory for memo state, resolved at most once per process.
+lisa_guard_memo_base=""
+
+# Whether a key was computed and a memo may be recorded on exit.
+lisa_guard_memo_ready=0
+
+# Digest tool for the memo key, resolved once. Empty disables the memo.
+lisa_guard_memo_digest_tool=""
+
+# Whether a directory is a real directory this user owns, and not a symlink —
+# builtins only, so this costs no process.
+#
+# `-O` is the load-bearing test: a memo planted by another user would suppress a
+# refusal, so a directory this user does not own is never used. The dispatcher
+# validates the whole parent chain before exporting `LISA_GUARD_MEMO_DIR`; this
+# re-checks the leaf rather than trusting the environment variable, because an
+# exported path is an input like any other.
+#
+# $1 - candidate directory
+lisa_guard_memo_dir_usable() {
+  [ -n "${1:-}" ] || return 1
+  [ -d "$1" ] || return 1
+  [ ! -L "$1" ] || return 1
+  [ -O "$1" ] || return 1
+}
+
+# Establish the memo base directory.
+#
+# Preferred source is `LISA_GUARD_MEMO_DIR`, exported by the dispatcher, which
+# has already resolved `TMPDIR` physically and walked every parent for ownership
+# and mode. Guards it spawns therefore pay nothing for that walk. A guard
+# reached by the plugin channel has no such parent, so it creates the directory
+# itself under `TMPDIR` — private by umask, and used only when this user owns
+# it. If neither works the memo is unavailable and every guard evaluates, which
+# is today's behaviour.
+lisa_guard_memo_resolve_base() {
+  [ -z "$lisa_guard_memo_base" ] || return 0
+
+  local candidate="${LISA_GUARD_MEMO_DIR:-}"
+  if lisa_guard_memo_dir_usable "$candidate"; then
+    lisa_guard_memo_base="$candidate"
+    return 0
+  fi
+
+  local uid="${LISA_GUARD_MEMO_UID:-}"
+  if [ -z "$uid" ]; then
+    uid="$(id -u 2>/dev/null || printf 'unknown')"
+  fi
+  candidate="${TMPDIR:-/tmp}"
+  candidate="${candidate%/}/lisa-guard-memo-$uid"
+  if [ ! -e "$candidate" ] && [ ! -L "$candidate" ]; then
+    (umask 077 && mkdir -p "$candidate") 2>/dev/null || true
+  fi
+  if lisa_guard_memo_dir_usable "$candidate"; then
+    lisa_guard_memo_base="$candidate"
+    return 0
+  fi
+  return 1
+}
+
+# Resolve a SHA-256 tool, or report that there is none.
+#
+# Two spellings because the guards run on macOS, on Linux CI images, and in
+# containers that carry neither. A missing digest tool disables dedupe rather
+# than degrading it to something weaker: a cheap checksum could collide, and a
+# collision here means a DIFFERENT payload reads as already-allowed, which is
+# lost enforcement rather than a slow guard.
+lisa_guard_memo_resolve_digest() {
+  [ -z "$lisa_guard_memo_digest_tool" ] || return 0
+  if command -v shasum >/dev/null 2>&1; then
+    lisa_guard_memo_digest_tool="shasum"
+  elif command -v sha256sum >/dev/null 2>&1; then
+    lisa_guard_memo_digest_tool="sha256sum"
+  else
+    return 1
+  fi
+  return 0
+}
+
+# Remove memo state older than a day.
+#
+# Files first, then the emptied session directories, both with `find -delete`.
+# Never a recursive forced delete: this runs unattended inside every guard, and
+# the blast radius of a mis-resolved base directory has to stay bounded to
+# entries this scheme itself created.
+lisa_guard_memo_sweep() {
+  find "$lisa_guard_memo_base" -mindepth 2 -maxdepth 2 -type f -mmin +1440 \
+    -delete 2>/dev/null || true
+  find "$lisa_guard_memo_base" -mindepth 1 -maxdepth 1 -type d -empty \
+    -mmin +1440 -delete 2>/dev/null || true
+}
+
+# Decide whether this guard has already been evaluated for this tool call, and
+# exit 0 in place of re-evaluating it when it has.
+#
+# Returns 0 for "carry on and evaluate". Never returns non-zero, so a caller
+# running under `set -e` is unaffected by it.
+#
+# $1 - guard name, used only to name the memo file
+# $2 - the raw tool payload this guard was handed on stdin
+lisa_guard_dedupe() {
+  local guard_name="${1:-}"
+  local payload="${2:-}"
+  [ -n "$guard_name" ] || return 0
+  [ -n "$payload" ] || return 0
+  [ -z "${LISA_GUARD_DEDUPE_DISABLE:-}" ] || return 0
+
+  # The session scopes every piece of state below. No session id means the
+  # payload came from a surface this cannot reason about — an adapter's
+  # synthesised envelope, a test harness — and the answer there is to evaluate.
+  local session_id=""
+  local session_pattern="\"session_id\"[[:space:]]*:[[:space:]]*\"([^\"]+)\""
+  if [[ "$payload" =~ $session_pattern ]]; then
+    session_id="${BASH_REMATCH[1]}"
+  fi
+  [ -n "$session_id" ] || return 0
+  # It is used as a path component, so anything that could escape the state
+  # directory disqualifies it outright.
+  case "$session_id" in *[!A-Za-z0-9._-]* | .*) return 0 ;; esac
+
+  lisa_guard_memo_resolve_base || return 0
+
+  local session_dir="$lisa_guard_memo_base/$session_id"
+  local fresh_session=0
+  if [ ! -d "$session_dir" ]; then
+    if (umask 077 && mkdir "$session_dir") 2>/dev/null; then
+      fresh_session=1
+    fi
+  fi
+  [ -d "$session_dir" ] || return 0
+  [ ! -L "$session_dir" ] || return 0
+
+  # Swept once, by the process that created the session directory, so a
+  # long-lived machine does not accumulate memo state. Off the hot path by
+  # construction: every later call in the session finds the directory present.
+  if [ "$fresh_session" -eq 1 ]; then
+    lisa_guard_memo_sweep
+  fi
+
+  # The directory this copy of the guard was served from. Two registrations of
+  # one guard are always two different directories — a repository tree or an
+  # applied `scripts/lisa-hooks/` on one side, a plugin cache on the other — so
+  # the set of directories seen this session IS the set of live channels.
+  local self_script="${BASH_SOURCE[1]:-$0}"
+  local self_dir="."
+  case "$self_script" in */*) self_dir="${self_script%/*}" ;; esac
+
+  local channels_file="$session_dir/channels"
+  local seen_self=0
+  local other_channel=0
+  local line=""
+  if [ -f "$channels_file" ]; then
+    while IFS= read -r line || [ -n "$line" ]; do
+      [ -n "$line" ] || continue
+      if [ "$line" = "$self_dir" ]; then
+        seen_self=1
+      else
+        other_channel=1
+      fi
+    done <"$channels_file"
+  fi
+  if [ "$seen_self" -eq 0 ]; then
+    printf '%s\n' "$self_dir" >>"$channels_file" 2>/dev/null || true
+  fi
+
+  # Single channel: the memo can never hit, so nothing below is worth its cost.
+  # This is the branch a host with no plugin takes on every call, and it is the
+  # reason the fork lives below this line rather than above it.
+  [ "$other_channel" -eq 1 ] || return 0
+
+  # The per-call discriminator. The transcript grows with every turn, so its
+  # size separates two identical commands issued at different points in one
+  # session — which matters because a guard classifies against a working tree
+  # that may have changed between them. Both channels read the same transcript
+  # within one tool call, and anything appended between them costs a memo miss,
+  # never a stale hit.
+  local transcript=""
+  local transcript_pattern="\"transcript_path\"[[:space:]]*:[[:space:]]*\"([^\"]+)\""
+  if [[ "$payload" =~ $transcript_pattern ]]; then
+    transcript="${BASH_REMATCH[1]}"
+  fi
+  [ -n "$transcript" ] || return 0
+  [ -f "$transcript" ] || return 0
+
+  lisa_guard_memo_resolve_digest || return 0
+
+  # The transcript's size, captured and VALIDATED before it can reach the
+  # digest, because a `stat` that rejects the other platform's spelling can
+  # still write to stdout on its way out. GNU `stat -f '%z' FILE` reads `-f` as
+  # `--file-system` and the format string as a FILE that does not exist: it
+  # exits non-zero AND prints a filesystem status block for the real operand,
+  # whose `Free` and `Available` counts move under any disk activity. Appended
+  # straight into the digest, that made the two channels of ONE tool call
+  # compute different keys, so the memo could only ever miss — the whole
+  # mechanism switched off on every Linux host, while the three "must not
+  # dedupe" controls stayed green because they assert a miss.
+  #
+  # Only an all-digits answer is accepted, and no size means NO MEMO rather
+  # than a memo with nothing separating one call from the next: that is
+  # condition 2 above, which the `A || B` spelling had left unenforced on a
+  # host carrying neither `stat`.
+  local transcript_size=""
+  transcript_size="$(stat -c '%s' "$transcript" 2>/dev/null)" ||
+    transcript_size=""
+  case "$transcript_size" in
+    "" | *[!0-9]*)
+      transcript_size="$(stat -f '%z' "$transcript" 2>/dev/null)" ||
+        transcript_size=""
+      ;;
+  esac
+  case "$transcript_size" in "" | *[!0-9]*) return 0 ;; esac
+
+  # One command substitution, covering everything the verdict depends on: the
+  # guard's own bytes, this library's bytes, the payload, and which call it is.
+  local key=""
+  key="$(
+    {
+      cat "$self_script" "${BASH_SOURCE[0]}" 2>/dev/null
+      printf '\n%s\n' "$payload"
+      printf '%s\n' "$transcript_size"
+    } | if [ "$lisa_guard_memo_digest_tool" = "shasum" ]; then
+      shasum -a 256 2>/dev/null
+    else
+      sha256sum 2>/dev/null
+    fi
+  )" || key=""
+  key="${key%% *}"
+  case "$key" in "" | *[!0-9a-f]*) return 0 ;; esac
+
+  # One memo file per guard per session, rewritten as the call moves on, so the
+  # state directory stays a fixed size however long a session runs. A saturated
+  # temp directory is its own outage on this fleet, and a memo scheme that grew
+  # a file per tool call would be one.
+  local memo_file="$session_dir/$guard_name.allow"
+  local recorded=""
+  if [ -f "$memo_file" ] && [ ! -L "$memo_file" ]; then
+    IFS= read -r recorded <"$memo_file" 2>/dev/null || recorded=""
+    if [ "$recorded" = "$key" ]; then
+      # A byte-identical copy of this guard already allowed this exact payload
+      # on this exact tool call. The evaluation being skipped could only have
+      # produced the verdict that was already delivered.
+      exit 0
+    fi
+  fi
+
+  lisa_guard_memo_key="$key"
+  lisa_guard_memo_file="$memo_file"
+  lisa_guard_memo_ready=1
+  return 0
+}
+
+# Record an ALLOW verdict for the other channel to replay. Installed as the
+# guard's EXIT trap, so it covers every path the guard can leave by.
+#
+# Any status other than 0 writes nothing and, when a memo from an earlier call
+# is still sitting there, removes it: a guard that refused must never leave
+# behind something that reads as permission.
+#
+# $1 - the exiting status
+lisa_guard_dedupe_record() {
+  [ "${lisa_guard_memo_ready:-0}" -eq 1 ] || return 0
+  [ -n "$lisa_guard_memo_file" ] || return 0
+  if [ "${1:-1}" -eq 0 ]; then
+    printf '%s\n' "$lisa_guard_memo_key" >"$lisa_guard_memo_file" 2>/dev/null ||
+      true
+  else
+    rm -f "$lisa_guard_memo_file" 2>/dev/null || true
+  fi
+  return 0
+}

package/plugins/lisa/hooks/inject-rules.sh

@@ -19,11 +19,19 @@ fi
 ROOT="${CLAUDE_PLUGIN_ROOT:-${PLUGIN_ROOT:-$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)}}"
 RULES_DIR="$ROOT/rules/eager"
 
-# Backward compatibility: if the eager subdir is absent (older Lisa install),
-# fall back to the flat rules/ directory so a partial upgrade still ships rules.
-if [ ! -d "$RULES_DIR" ]; then
-  RULES_DIR="$ROOT/rules"
-fi
+# There is deliberately NO fallback to a flat rules/ directory (#3993).
+#
+# It existed to cover an "older Lisa install", but a plugin's rules and this
+# script ship in the same directory and install as one unit, so that skew cannot
+# arise. What the branch actually did was absorb plugins that never adopted the
+# split: `lisa-phaser` shipped a flat 11,007-byte rules/phaser.md on 2026-06-11,
+# two weeks AFTER the split commit a820527 (2026-05-28), and the fallback
+# injected the whole body at every SessionStart and SubagentStart without any
+# surface reporting it. A quiet success is indistinguishable from a tree that was
+# never split. With the branch gone, an unsplit plugin injects nothing and gets
+# noticed. (The Codex mirror injector keeps its fallback: `.codex/lisa-rules/` is
+# written separately from the script that reads it, so vintage skew IS possible
+# there.)
 
 # Bail silently if no rules directory at all
 [ -d "$RULES_DIR" ] || exit 0