switchroom 0.21.15 → 0.21.16

package/bin/rules-sentinel-hook.sh

@@ -0,0 +1,101 @@
+#!/bin/bash
+# SessionStart hook — Memory v2 M1 rules-block tamper sentinel
+# (carve-M1.md §3, red-team-M1.md §B/§C, MINOR fixes #6/#7).
+#
+# Recomputes the rules block's sha256 sentinel + the mutation-log hash
+# chain and compares against what's embedded in the agent's own
+# CLAUDE.md. Purely advisory in M1 (dark build — no `memory.rules_block`
+# agent exists live yet): this hook exists so the wiring, the read-only
+# discipline, and the fast no-op path are all proven out and tested
+# BEFORE M3 flips the flag on any real agent.
+#
+# STRICT DISCIPLINE (red-team §C, root-owned-overlay EACCES hazard):
+#   - READ-ONLY. This script must NEVER write to the agent's tree. A
+#     root-invoked write here would leave a root-owned file behind and
+#     EACCES the agent's own subsequent writes to it (the exact failure
+#     class that dropped clerk's crons for weeks — see this repo's
+#     CLAUDE.md "root-tier host access" section). `switchroom memory
+#     rule verify` (the CLI path) is the only writer-adjacent tool, and
+#     even IT never writes — verify is read-only too.
+#   - Prints to STDOUT, not stderr. SessionStart hook stdout is added to
+#     the model's context (unlike PreCompact); a stderr message would be
+#     silently dropped by Claude Code's hook contract, defeating the
+#     entire point of a tamper *sentinel* — see this file's own header
+#     precedent in bin/timezone-hook.sh for the additionalContext
+#     mechanism this hook reuses (NOT because "no additionalContext
+#     emitter exists" — one already does, in timezone-hook.sh and
+#     scaffold.ts; this hook simply needs the same JSON shape).
+#   - No-ops FAST and SILENTLY when the rules-block markers are absent
+#     (the M1/M3-flag-off dark case, and the pre-first-rule case even
+#     with the flag on) — never print a divergence notice about a block
+#     that was never created.
+#
+# Wiring note (MINOR fix #7): unlike working-state-reload-hook.sh
+# (matcher: "compact" — fires ONLY on compaction), this hook is wired
+# with NO matcher, so it fires on every SessionStart source (startup,
+# resume, clear, fork, compact). That is a deliberate divergence: a
+# tamper sentinel that only checked itself after compaction would miss
+# tampering that happened between a fresh boot and the first compaction
+# (the common case for a short-lived session). The check itself is a
+# single local file read + a hash compare — cheap enough to run on every
+# boot without a latency concern (contrast working-state-reload's
+# network-hop briefing, which specifically justified restricting itself
+# to the compact matcher).
+#
+# Exit code: 0 always. A hook that exits non-zero BLOCKS the turn in
+# Claude Code; a tamper signal must degrade to a visible notice, never
+# a hard stop (the same "failure modes are silent" doctrine as
+# timezone-hook.sh).
+
+set -u
+
+CLAUDE_MD="${CLAUDE_PROJECT_DIR:-}/CLAUDE.md"
+
+if [ -z "${CLAUDE_PROJECT_DIR:-}" ] || [ ! -f "$CLAUDE_MD" ]; then
+  exit 0
+fi
+
+# Fast no-op: markers absent (dark build / no rules yet). grep -q is a
+# single pass over the file, no subshell fork per line.
+if ! grep -q '<!-- switchroom:rules:begin -->' "$CLAUDE_MD" 2>/dev/null; then
+  exit 0
+fi
+
+# Delegate the actual sentinel recompute to the CLI verb this module
+# already ships (`switchroom memory rule verify <agent>`), which reuses
+# rules-store.ts's verifyIntegrity — one recompute implementation, not
+# two (a bash reimplementation of the sha256-over-canonical-JSON
+# encoding would be a second place for the two encodings to drift).
+# `switchroom` is expected on PATH inside the agent container image;
+# if it's missing, degrade silently rather than error the turn.
+if ! command -v switchroom >/dev/null 2>&1; then
+  exit 0
+fi
+
+AGENT_NAME="$(basename "${CLAUDE_PROJECT_DIR:-}")"
+VERIFY_OUT="$(switchroom memory rule verify "$AGENT_NAME" 2>&1)"
+VERIFY_EXIT=$?
+
+# Exit-code contract (memory-rules.ts `verify`):
+#   0 = clean; 2 = GENUINE TAMPER; any other non-zero = an ENVIRONMENT
+#   failure (config unreadable, agent-dir mismatch, withConfigError exit 1,
+#   `switchroom` bug). Only code 2 warrants injecting a tamper notice into
+#   the model's context — an environment failure that hard-injected "tamper
+#   FAILED" would cry wolf on every broken boot and erode the signal. Degrade
+#   silently on everything that is not an unambiguous tamper.
+if [ "$VERIFY_EXIT" -ne 2 ]; then
+  exit 0
+fi
+
+MSG="Rules-block tamper sentinel FAILED for \"$AGENT_NAME\": ${VERIFY_OUT}"
+
+if command -v jq >/dev/null 2>&1; then
+  jq -cn --arg msg "$MSG" \
+    '{hookSpecificOutput:{hookEventName:"SessionStart",additionalContext:$msg}}'
+else
+  ESCAPED=${MSG//\\/\\\\}
+  ESCAPED=${ESCAPED//\"/\\\"}
+  printf '{"hookSpecificOutput":{"hookEventName":"SessionStart","additionalContext":"%s"}}\n' "$ESCAPED"
+fi
+
+exit 0

package/dist/agent-scheduler/index.js

@@ -11118,6 +11118,7 @@ var AgentMemorySchema = exports_external.object({
     }
   }).optional().describe("Operator-declared, per-specialist Hindsight mental models (RFC " + "Phase 5). Named, opt-in curated reflections this agent's bank should " + "carry — e.g. a coach's 'training-plan-state' or a lawyer's " + "'open-matters'. Ensured idempotently at scaffold/reconcile: NOTHING " + "is created unless declared here (zero declarations = zero models, " + "matching post-#2447 behaviour), and no fixed identity model is " + "reintroduced — 'who the user is' stays owned by dedicated profile " + "banks (users.*.profile_bank), never a per-agent model. Per-agent " + "ONLY: intentionally not accepted at the defaults/profile tier, so a " + "model can never be fleet-seeded — each specialist opts in on its own " + "(the invariant-clean inverse of the retired blind auto-seeding)."),
   observations_mission: exports_external.string().optional().describe("Steers what the observation-consolidation LLM synthesises from raw " + "facts (the higher-order 'what patterns matter' lens). Cascade: override."),
+  rules_block: exports_external.boolean().default(false).describe("Memory v2 M3 go-live flag (M1 only DEFINES this, default false; " + "M3 flips it per agent). When true, the sanctioned rules/index " + "blocks render live in this agent's CLAUDE.md (marker-delimited, " + "below the `# --- Yours ---` line) AND the permission deny for " + "direct Edit/Write of the agent's own CLAUDE.md is seeded — the " + "flag couples deny + tools together so a live deny never orphans " + "the invited free-text edit path (the `memory_edit_yours` verb is " + "the only sanctioned writer once flipped). Unset/false ⇒ byte-" + "identical to pre-M1 behaviour: no blocks, no deny, dark build. " + "Cascade: override (per-agent wins over default; never fleet-" + "seeded — each agent's flip is a deliberate M3 rollout step)."),
   disposition: exports_external.object({
     skepticism: exports_external.number().int().min(1).max(5).optional().describe("How much the bank doubts unverified claims (1-5; engine default 3)."),
     literalism: exports_external.number().int().min(1).max(5).optional().describe("How literally the bank reads statements vs inferring intent (1-5; engine default 3)."),

package/dist/auth-broker/index.js

@@ -11143,6 +11143,7 @@ var init_schema = __esm(() => {
       }
     }).optional().describe("Operator-declared, per-specialist Hindsight mental models (RFC " + "Phase 5). Named, opt-in curated reflections this agent's bank should " + "carry — e.g. a coach's 'training-plan-state' or a lawyer's " + "'open-matters'. Ensured idempotently at scaffold/reconcile: NOTHING " + "is created unless declared here (zero declarations = zero models, " + "matching post-#2447 behaviour), and no fixed identity model is " + "reintroduced — 'who the user is' stays owned by dedicated profile " + "banks (users.*.profile_bank), never a per-agent model. Per-agent " + "ONLY: intentionally not accepted at the defaults/profile tier, so a " + "model can never be fleet-seeded — each specialist opts in on its own " + "(the invariant-clean inverse of the retired blind auto-seeding)."),
     observations_mission: exports_external.string().optional().describe("Steers what the observation-consolidation LLM synthesises from raw " + "facts (the higher-order 'what patterns matter' lens). Cascade: override."),
+    rules_block: exports_external.boolean().default(false).describe("Memory v2 M3 go-live flag (M1 only DEFINES this, default false; " + "M3 flips it per agent). When true, the sanctioned rules/index " + "blocks render live in this agent's CLAUDE.md (marker-delimited, " + "below the `# --- Yours ---` line) AND the permission deny for " + "direct Edit/Write of the agent's own CLAUDE.md is seeded — the " + "flag couples deny + tools together so a live deny never orphans " + "the invited free-text edit path (the `memory_edit_yours` verb is " + "the only sanctioned writer once flipped). Unset/false ⇒ byte-" + "identical to pre-M1 behaviour: no blocks, no deny, dark build. " + "Cascade: override (per-agent wins over default; never fleet-" + "seeded — each agent's flip is a deliberate M3 rollout step)."),
     disposition: exports_external.object({
       skepticism: exports_external.number().int().min(1).max(5).optional().describe("How much the bank doubts unverified claims (1-5; engine default 3)."),
       literalism: exports_external.number().int().min(1).max(5).optional().describe("How literally the bank reads statements vs inferring intent (1-5; engine default 3)."),

package/dist/cli/notion-write-pretool.mjs

@@ -11879,6 +11879,7 @@ var AgentMemorySchema = exports_external.object({
     }
   }).optional().describe("Operator-declared, per-specialist Hindsight mental models (RFC " + "Phase 5). Named, opt-in curated reflections this agent's bank should " + "carry \u2014 e.g. a coach's 'training-plan-state' or a lawyer's " + "'open-matters'. Ensured idempotently at scaffold/reconcile: NOTHING " + "is created unless declared here (zero declarations = zero models, " + "matching post-#2447 behaviour), and no fixed identity model is " + "reintroduced \u2014 'who the user is' stays owned by dedicated profile " + "banks (users.*.profile_bank), never a per-agent model. Per-agent " + "ONLY: intentionally not accepted at the defaults/profile tier, so a " + "model can never be fleet-seeded \u2014 each specialist opts in on its own " + "(the invariant-clean inverse of the retired blind auto-seeding)."),
   observations_mission: exports_external.string().optional().describe("Steers what the observation-consolidation LLM synthesises from raw " + "facts (the higher-order 'what patterns matter' lens). Cascade: override."),
+  rules_block: exports_external.boolean().default(false).describe("Memory v2 M3 go-live flag (M1 only DEFINES this, default false; " + "M3 flips it per agent). When true, the sanctioned rules/index " + "blocks render live in this agent's CLAUDE.md (marker-delimited, " + "below the `# --- Yours ---` line) AND the permission deny for " + "direct Edit/Write of the agent's own CLAUDE.md is seeded \u2014 the " + "flag couples deny + tools together so a live deny never orphans " + "the invited free-text edit path (the `memory_edit_yours` verb is " + "the only sanctioned writer once flipped). Unset/false \u21d2 byte-" + "identical to pre-M1 behaviour: no blocks, no deny, dark build. " + "Cascade: override (per-agent wins over default; never fleet-" + "seeded \u2014 each agent's flip is a deliberate M3 rollout step)."),
   disposition: exports_external.object({
     skepticism: exports_external.number().int().min(1).max(5).optional().describe("How much the bank doubts unverified claims (1-5; engine default 3)."),
     literalism: exports_external.number().int().min(1).max(5).optional().describe("How literally the bank reads statements vs inferring intent (1-5; engine default 3)."),