From 2fc523491122c23229f8034ff054cd1c14cbc02d Mon Sep 17 00:00:00 2001
From: Boden Fuller <boden.fuller@gmail.com>
Date: Thu, 2 Apr 2026 15:52:33 -0400
Subject: [PATCH] refactor(hooks): keep Claude startup surface in CLAUDE.md

---
 CLAUDE.md                                     |  11 +
 README.md                                     |   9 +-
 cli/README.md                                 |  13 +-
 cli/embedded/hooks/factory-router.sh          |  51 +---
 cli/embedded/hooks/session-start.sh           | 218 +-----------------
 docs/ENV-VARS.md                              |   6 +-
 docs/SKILLS.md                                |   2 +-
 docs/cli-skills-map.md                        |   2 +-
 docs/contracts/hook-runtime-contract.md       |   6 +-
 .../session-intelligence-trust-model.md       |   4 +-
 docs/how-it-works.md                          |   4 +-
 docs/workflows/session-lifecycle.md           |  14 +-
 hooks/factory-router.sh                       |  51 +---
 hooks/session-start.sh                        | 218 +-----------------
 tests/hooks/hook-output-schema.bats           |  24 +-
 tests/hooks/test-hook-lifecycle.sh            |  19 +-
 tests/hooks/test-hooks.bats                   |  20 +-
 tests/hooks/test-hooks.sh                     |  80 +++----
 tests/integration/test-hook-chain.sh          |  35 +--
 19 files changed, 111 insertions(+), 676 deletions(-)

diff --git a/CLAUDE.md b/CLAUDE.md
index 4ff184611..3363e1cf6 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -194,6 +194,17 @@ Research → Plan → Implement → Validate
     └──── Knowledge Flywheel ────┘
 ```
 
+## Claude Code Startup Surface
+
+`CLAUDE.md` is the startup surface in Claude Code. Do not expect `SessionStart`
+or first-prompt hooks to inject briefings into the conversation.
+
+- Use the goal stated in the user prompt or recovered handoff as the working objective.
+- If you want the full software-factory lane, run `/rpi "goal"` explicitly.
+- If you want a compiled goal-time briefing first, run `ao knowledge brief --goal "goal"`.
+- Treat `.agents/ao/factory-goal.txt` and `.agents/ao/factory-briefing.txt` as
+  silent runtime state, not operator-facing instructions.
+
 ### Session & Swarm Constraints
 
 - **Multi-phase work:** Route through `ao rpi` — it enforces 90-min phase timeouts and 10-min stall detection. Raw sessions have neither.
diff --git a/README.md b/README.md
index cc78e117f..a70e23a4a 100644
--- a/README.md
+++ b/README.md
@@ -78,11 +78,10 @@ Or install via [release binaries](https://github.com/boshu2/agentops/releases) o
 
 Then type `/quickstart` in your agent chat.
 
-In Claude Code, the native hook path now behaves like a software-factory
-startup surface: `SessionStart` prefers a matched goal-time briefing when
-handoff or tracker state already gives the runtime a goal, and when no goal is
-recovered the first substantive prompt becomes intake and the plugin routes the
-session toward `/rpi`.
+In Claude Code, `CLAUDE.md` is the startup surface. Installed hooks stay
+silent: `SessionStart` prepares runtime state and can stage factory goal or
+briefing files, while `UserPromptSubmit` can capture first-prompt intake
+without injecting additional context into the session.
 
 ---
 
diff --git a/cli/README.md b/cli/README.md
index 7ef55c1e3..fccb4fdec 100644
--- a/cli/README.md
+++ b/cli/README.md
@@ -18,19 +18,16 @@ ao init --hooks
 ao hooks test
 ```
 
-That's it. In Claude Code, the installed hooks now provide the native factory
-startup lane: `SessionStart` prefers matched knowledge briefings, and the first
-substantive prompt can become intake when no startup goal exists.
+That's it. In Claude Code, `CLAUDE.md` remains the startup surface. The
+installed hooks stay silent and only prepare runtime state for the factory lane.
 
 ## What It Does
 
-**SessionStart**: Performs startup maintenance, then surfaces the best
-available factory startup context. If a goal is already known, it prefers a
-matched knowledge briefing and treats ranked learnings as supporting evidence.
+**SessionStart**: Performs startup maintenance, recovers handoff state, and can
+stage `factory-goal.txt` / `factory-briefing.txt` without injecting context.
 
 **UserPromptSubmit**: When startup lacked a goal, the first substantive prompt
-is captured as factory intake, a goal-time briefing is built if possible, and
-the runtime is nudged toward `/rpi`.
+can be captured as silent factory intake and staged for later explicit use.
 
 **SessionEnd**: Extracts learnings and updates the feedback loop.
 
diff --git a/cli/embedded/hooks/factory-router.sh b/cli/embedded/hooks/factory-router.sh
index 3407e378e..a3ee5b7ff 100755
--- a/cli/embedded/hooks/factory-router.sh
+++ b/cli/embedded/hooks/factory-router.sh
@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
-# factory-router.sh - UserPromptSubmit hook: convert first substantive prompt into
-# a goal-time briefing when SessionStart had no factory objective.
+# factory-router.sh - UserPromptSubmit hook: capture first substantive prompt as
+# factory state when SessionStart had no goal. This hook stays silent.
 
 [ "${AGENTOPS_HOOKS_DISABLED:-}" = "1" ] && exit 0
 [ "${AGENTOPS_FACTORY_ROUTER_DISABLED:-}" = "1" ] && exit 0
@@ -64,26 +64,6 @@ build_factory_briefing() {
     printf '%s' "$path"
 }
 
-read_factory_briefing() {
-    local path="$1"
-    local max_chars=2200
-    local content trimmed
-
-    [ -f "$path" ] || return 0
-    content=$(cat "$path" 2>/dev/null || true)
-    [ -n "$content" ] || return 0
-
-    if [ "${#content}" -gt "$max_chars" ]; then
-        trimmed="${content:0:$max_chars}"
-        content="${trimmed%
-*}
-
-*[briefing truncated by factory router]*"
-    fi
-
-    printf '%s' "$content"
-}
-
 extract_prompt() {
     local input="$1"
     if command -v jq >/dev/null 2>&1; then
@@ -157,37 +137,10 @@ printf '%s' "$GOAL" > "$GOAL_FILE" 2>/dev/null || true
 rm -f "$INTAKE_FLAG" 2>/dev/null || true
 printf '%s\n' "${CLAUDE_SESSION_ID:-unknown}" > "$ROUTER_FLAG" 2>/dev/null || true
 
-MESSAGE="Factory intake captured from the first substantive prompt.
-- **Goal:** ${GOAL}"
-
 if [ -n "$BRIEFING_PATH" ]; then
     printf '%s' "$BRIEFING_PATH" > "$BRIEFING_FILE" 2>/dev/null || true
-    BRIEFING_CONTENT="$(read_factory_briefing "$BRIEFING_PATH")"
-    MESSAGE="${MESSAGE}
-- **Primary briefing:** ${BRIEFING_PATH}
-- **Delivery lane:** Continue naturally, or run \`/rpi \"${GOAL}\"\`.
-
-Treat the briefing below as the primary dynamic surface for this session.
-
-<FACTORY_BRIEFING>
-${BRIEFING_CONTENT}
-</FACTORY_BRIEFING>"
 else
     rm -f "$BRIEFING_FILE" 2>/dev/null || true
-    MESSAGE="${MESSAGE}
-- **Primary briefing:** not available yet
-- **Delivery lane:** Run \`/rpi \"${GOAL}\"\` or continue with this objective.
-
-No matched knowledge briefing was available for the intake goal, so the captured goal above is the active factory objective."
-fi
-
-if command -v jq >/dev/null 2>&1; then
-    jq -n --arg msg "$MESSAGE" '{"hookSpecificOutput":{"hookEventName":"UserPromptSubmit","additionalContext":$msg}}'
-else
-    safe_msg=${MESSAGE//\\/\\\\}
-    safe_msg=${safe_msg//\"/\\\"}
-    safe_msg=$(printf '%s' "$safe_msg" | tr '\n' ' ')
-    echo "{\"hookSpecificOutput\":{\"hookEventName\":\"UserPromptSubmit\",\"additionalContext\":\"$safe_msg\"}}"
 fi
 
 exit 0
diff --git a/cli/embedded/hooks/session-start.sh b/cli/embedded/hooks/session-start.sh
index f47a97dec..5953a6888 100755
--- a/cli/embedded/hooks/session-start.sh
+++ b/cli/embedded/hooks/session-start.sh
@@ -1,8 +1,7 @@
 #!/usr/bin/env bash
 # AgentOps Session Start Hook
-# Creates .agents/ directories, consumes handoffs, and surfaces startup context.
-# In default factory mode, SessionStart prefers a matched knowledge briefing when
-# a goal is available; manual mode keeps a lighter startup surface.
+# Creates .agents/ directories, consumes handoffs, and prepares runtime state.
+# CLAUDE.md owns the operator-facing startup surface; this hook stays silent.
 
 # Kill switches
 [ "${AGENTOPS_HOOKS_DISABLED:-}" = "1" ] && exit 0
@@ -83,13 +82,6 @@ derive_lookup_query() {
     return 0
 }
 
-detect_lookup_bead() {
-    if ! command -v bd >/dev/null 2>&1; then
-        return 0
-    fi
-    run_with_timeout 1 bd current | head -1 | tr -d '\r' | sed 's/^ //; s/ $//'
-}
-
 build_factory_briefing() {
     local goal="$1"
     local output path
@@ -108,26 +100,6 @@ build_factory_briefing() {
     printf '%s' "$path"
 }
 
-read_factory_briefing() {
-    local path="$1"
-    local max_chars=2200
-    local content trimmed
-
-    [ -f "$path" ] || return 0
-    content=$(cat "$path" 2>/dev/null || true)
-    [ -n "$content" ] || return 0
-
-    if [ "${#content}" -gt "$max_chars" ]; then
-        trimmed="${content:0:$max_chars}"
-        content="${trimmed%
-*}
-
-*[briefing truncated by session-start hook]*"
-    fi
-
-    printf '%s' "$content"
-}
-
 write_environment_manifest() {
     local env_file="$AO_DIR/environment.json"
     local tmp_file git_branch head_sha git_dirty tools_json manifest_json
@@ -300,27 +272,11 @@ rpi/*
 EOF
 fi
 
-# Flywheel behavior
-INJECTED_KNOWLEDGE=""
 STARTUP_CONTEXT_MODE="$(resolve_startup_context_mode)"
 FACTORY_GOAL=""
 FACTORY_BRIEFING_PATH=""
-FACTORY_BRIEFING_CONTENT=""
-
-# Derive MEMORY.md path once (used for stale nudge)
-MEMORY_DIR="$HOME/.claude/projects"
-PROJECT_PATH=$(printf '%s' "$ROOT" | tr '/' '-')
-MEMORY_FILE="$MEMORY_DIR/$PROJECT_PATH/memory/MEMORY.md"
-MEMORY_AGE_DAYS=-1
-if [ -f "$MEMORY_FILE" ]; then
-    MTIME=$(stat -f %m "$MEMORY_FILE" 2>/dev/null || stat -c %Y "$MEMORY_FILE" 2>/dev/null || echo "")
-    if [ -n "$MTIME" ]; then
-        MEMORY_AGE_DAYS=$(( ($(date +%s) - MTIME) / 86400 ))
-    fi
-fi
 
 # Structured handoff consumption (ao handoff JSON artifacts)
-HANDOFF_CONTEXT=""
 if [ -d "$ROOT/.agents/handoff" ] && command -v jq &>/dev/null; then
     # Find newest unconsumed .json handoff (exclude .consumed.json and .consuming.json)
     HANDOFF_JSON=$(find "$ROOT/.agents/handoff" -maxdepth 1 -name 'handoff-*.json' \
@@ -340,52 +296,10 @@ if [ -d "$ROOT/.agents/handoff" ] && command -v jq &>/dev/null; then
                 "$CONSUMING" > "${CONSUMING}.tmp" 2>/dev/null \
                 && mv "${CONSUMING}.tmp" "${HANDOFF_JSON%.json}.consumed.json" 2>/dev/null \
                 && rm -f "$CONSUMING" 2>/dev/null
-            # Build injection context
-            HANDOFF_CONTEXT="### Handoff Context (${H_TYPE})"
-            [ -n "$H_GOAL" ] && HANDOFF_CONTEXT="${HANDOFF_CONTEXT}
-- **Goal:** ${H_GOAL}"
-            [ -n "$H_SUMMARY" ] && HANDOFF_CONTEXT="${HANDOFF_CONTEXT}
-- **Summary:** ${H_SUMMARY}"
-            [ -n "$H_CONTINUATION" ] && HANDOFF_CONTEXT="${HANDOFF_CONTEXT}
-- **Continue:** ${H_CONTINUATION}"
-            HANDOFF_CONTEXT="${HANDOFF_CONTEXT}
-- **Source:** ${HANDOFF_JSON}"
         fi
     fi
 fi
 
-# Predecessor handoff discovery
-PREDECESSOR_FILE="${GT_PREDECESSOR_HANDOFF:-}"
-if [ -z "$PREDECESSOR_FILE" ] && [ -d "$ROOT/.agents/handoff" ]; then
-    PREDECESSOR_FILE=$(ls -t "$ROOT/.agents/handoff/"*.md 2>/dev/null | head -1)
-fi
-
-# Build injection context (MEMORY.md is auto-loaded by Claude Code)
-INJECTED_KNOWLEDGE="MEMORY.md is auto-loaded by Claude Code for this project.
-For on-demand retrieval: \`ao search \"<query>\"\` or \`ao lookup --query \"<query>\"\`"
-if [ -n "$HANDOFF_CONTEXT" ]; then
-    INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-${HANDOFF_CONTEXT}"
-fi
-if [ -n "$PREDECESSOR_FILE" ] && [ -f "$PREDECESSOR_FILE" ]; then
-    INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-Predecessor handoff: ${PREDECESSOR_FILE}"
-fi
-
-# Point to knowledge signpost (agents search on demand)
-if [ -f "${ROOT:-.}/.agents/AGENTS.md" ]; then
-    INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-Knowledge artifacts are in \`.agents/\`. See \`.agents/AGENTS.md\` for navigation.
-Use \`ao lookup --query \"topic\"\` for learnings retrieval, or \`Grep\` in \`.agents/learnings/\` as fallback."
-else
-    INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-Knowledge artifacts are in \`.agents/\` (if populated).
-Use \`ao lookup --query \"topic\"\` for on-demand learnings retrieval."
-fi
-
 if [ "$STARTUP_CONTEXT_MODE" = "factory" ]; then
     FACTORY_GOAL="$(derive_lookup_query)"
     if [ -n "$FACTORY_GOAL" ]; then
@@ -393,138 +307,12 @@ if [ "$STARTUP_CONTEXT_MODE" = "factory" ]; then
         FACTORY_BRIEFING_PATH="$(build_factory_briefing "$FACTORY_GOAL")"
         if [ -n "$FACTORY_BRIEFING_PATH" ]; then
             printf '%s' "$FACTORY_BRIEFING_PATH" > "$ROOT/.agents/ao/factory-briefing.txt" 2>/dev/null || true
-            FACTORY_BRIEFING_CONTENT="$(read_factory_briefing "$FACTORY_BRIEFING_PATH")"
-            INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-### Factory Startup Surface
-- **Goal:** ${FACTORY_GOAL}
-- **Primary briefing:** ${FACTORY_BRIEFING_PATH}
-- **Delivery lane:** Continue naturally, or run \`/rpi \"${FACTORY_GOAL}\"\` for the full factory cycle.
-
-Treat the matched briefing below as the primary dynamic startup surface for this session. Use ranked learnings only as supporting evidence.
-
-<FACTORY_BRIEFING>
-${FACTORY_BRIEFING_CONTENT}
-</FACTORY_BRIEFING>"
         else
-            INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-### Factory Startup Surface
-- **Goal:** ${FACTORY_GOAL}
-- **Primary briefing:** not available yet
-- **Delivery lane:** Run \`/rpi \"${FACTORY_GOAL}\"\` or continue with this objective.
-
-No matched knowledge briefing was available at session start. Treat the goal above as the active factory objective; supporting learnings below are advisory rather than the primary brief."
+            rm -f "$ROOT/.agents/ao/factory-briefing.txt" 2>/dev/null || true
         fi
     else
         : > "$ROOT/.agents/ao/.factory-intake-needed" 2>/dev/null || true
-        INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-### Factory Startup Surface
-No startup goal was recovered from handoff or tracker state.
-Treat the first substantive user prompt as factory intake: compile a goal-time briefing, then route into \`/rpi\`."
     fi
 fi
 
-# Auto-retrieve and cite top learnings for this session (closes citation gap)
-if command -v ao &>/dev/null; then
-    LOOKUP_QUERY="$(derive_lookup_query)"
-    LOOKUP_BEAD="$(detect_lookup_bead)"
-    LOOKUP_CONTEXT=""
-    LOOKUP_ARGS=(lookup --limit 5)
-    if [ -n "$LOOKUP_QUERY" ]; then
-        LOOKUP_ARGS+=(--query "$LOOKUP_QUERY")
-        LOOKUP_CONTEXT="query=\"$LOOKUP_QUERY\""
-    fi
-    if [ -n "$LOOKUP_BEAD" ]; then
-        LOOKUP_ARGS+=(--bead "$LOOKUP_BEAD")
-        if [ -n "$LOOKUP_CONTEXT" ]; then
-            LOOKUP_CONTEXT="${LOOKUP_CONTEXT}, "
-        fi
-        LOOKUP_CONTEXT="${LOOKUP_CONTEXT}bead=${LOOKUP_BEAD}"
-    fi
-    if [ -z "$LOOKUP_CONTEXT" ]; then
-        LOOKUP_CONTEXT="task context unavailable"
-    fi
-    if [ -n "$LOOKUP_QUERY" ] || [ -n "$LOOKUP_BEAD" ]; then
-        FLYWHEEL_KNOWLEDGE=$(run_with_timeout 5 ao "${LOOKUP_ARGS[@]}" || true)
-    else
-        FLYWHEEL_KNOWLEDGE=""
-    fi
-    if [ -n "$FLYWHEEL_KNOWLEDGE" ]; then
-        INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-### Supporting Learnings (auto-retrieved: ${LOOKUP_CONTEXT})
-${FLYWHEEL_KNOWLEDGE}"
-    fi
-fi
-
-# Keep startup context lean: inject only fresh flywheel knowledge and a short skill pointer.
-SKILL_FILE="${PLUGIN_ROOT}/skills/using-agentops/SKILL.md"
-if [ -f "$SKILL_FILE" ]; then
-    using_agentops_hint="AgentOps workflow context is available. Use the Skill tool to read ${SKILL_FILE} when needed."
-else
-    using_agentops_hint="(AgentOps skill content unavailable at ${SKILL_FILE})"
-fi
-
-# Truncation (static cap — no notebook mode logic needed)
-MAX_INJECT_CHARS=4000
-if [ -n "$INJECTED_KNOWLEDGE" ] && [ "${#INJECTED_KNOWLEDGE}" -gt "$MAX_INJECT_CHARS" ]; then
-    # Truncate at last newline within budget (never mid-line)
-    trimmed="${INJECTED_KNOWLEDGE:0:$MAX_INJECT_CHARS}"
-    INJECTED_KNOWLEDGE="${trimmed%
-*}
-
-*[truncated by session-start hook]*"
-fi
-
-# Nudge agent if MEMORY.md is stale or missing
-if [ -n "${MEMORY_FILE:-}" ] && [ -f "$MEMORY_FILE" ] && [ "${MEMORY_AGE_DAYS:-0}" -gt 7 ]; then
-    INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-*Note: Your MEMORY.md hasn't been updated in ${MEMORY_AGE_DAYS} days. Consider running \`ao notebook update\` or updating it manually.*"
-fi
-
-if [ -n "$INJECTED_KNOWLEDGE" ]; then
-    full_content=$(cat <<HOOKCTX
-<AGENTOPS_CONTEXT>
-${INJECTED_KNOWLEDGE}
-
-${using_agentops_hint}
-</AGENTOPS_CONTEXT>
-HOOKCTX
-)
-else
-    full_content=$(cat <<HOOKCTX
-<AGENTOPS_CONTEXT>
-No prior flywheel knowledge was injected for this session.
-${using_agentops_hint}
-</AGENTOPS_CONTEXT>
-HOOKCTX
-)
-fi
-
-if command -v jq &>/dev/null; then
-    additional_context=$(printf '%s' "$full_content" | jq -Rs '.')
-    cat <<HOOKEOF
-{
-  "hookSpecificOutput": {
-    "hookEventName": "SessionStart",
-    "additionalContext": ${additional_context}
-  }
-}
-HOOKEOF
-else
-    # Minimal fallback — escape newlines and quotes
-    escaped=$(printf '%s' "$full_content" | sed 's/\\/\\\\/g; s/"/\\"/g' | tr '\n' ' ')
-    cat <<HOOKEOF
-{
-  "hookSpecificOutput": {
-    "hookEventName": "SessionStart",
-    "additionalContext": "${escaped}"
-  }
-}
-HOOKEOF
-fi
-
 exit 0
diff --git a/docs/ENV-VARS.md b/docs/ENV-VARS.md
index 3216bfe77..e5835eae7 100644
--- a/docs/ENV-VARS.md
+++ b/docs/ENV-VARS.md
@@ -59,9 +59,9 @@ These control the optional hook system installed via `ao init --hooks`. Each hoo
 |----------|---------|-------------|
 | `AGENTOPS_HOOKS_DISABLED` | `0` | Set to `1` to disable **all** hooks at once. Global kill switch. Useful for debugging or when hooks interfere with a specific workflow. |
 | `AGENTOPS_SESSION_START_DISABLED` | `0` | Set to `1` to disable the session-start hook entirely. |
-| `AGENTOPS_STARTUP_CONTEXT_MODE` | `factory` | Controls SessionStart behavior: `factory` (default — prefer matched goal-time briefings, keep ranked learnings as supporting context, and mark missing-goal sessions for prompt-time intake) or `manual` (lighter startup surface without factory briefing/routing). Legacy values `lean` and `legacy` are treated as `manual` for compatibility. |
-| `AGENTOPS_STARTUP_LEGACY_INJECT` | (unset) | Compatibility alias that forces `manual` startup mode regardless of `AGENTOPS_STARTUP_CONTEXT_MODE`. Use only as a rollback for older startup expectations. |
-| `AGENTOPS_FACTORY_ROUTER_DISABLED` | `0` | Set to `1` to disable the `UserPromptSubmit` factory intake hook that captures the first substantive prompt and routes the session toward `/rpi` when startup had no goal. |
+| `AGENTOPS_STARTUP_CONTEXT_MODE` | `factory` | Controls SessionStart behavior: `factory` (default — silently stage factory goal/briefing state when a goal is recoverable, and mark missing-goal sessions for prompt-time intake) or `manual` (baseline startup maintenance only). Legacy values `lean` and `legacy` are treated as `manual` for compatibility. |
+| `AGENTOPS_STARTUP_LEGACY_INJECT` | (unset) | Compatibility alias that forces `manual` startup mode regardless of `AGENTOPS_STARTUP_CONTEXT_MODE`. Use only as a rollback for older startup expectations that relied on hidden hook-managed startup context. |
+| `AGENTOPS_FACTORY_ROUTER_DISABLED` | `0` | Set to `1` to disable the `UserPromptSubmit` factory intake hook that captures the first substantive prompt into `.agents/ao/` state when startup had no goal. |
 | `AGENTOPS_MEMORY_SYNC` | `0` | Set to `1` to enable repo-root MEMORY.md sync at SessionEnd. Writes session history with managed block markers for cross-runtime access (Codex, OpenCode). |
 | `AGENTOPS_EVICTION_DISABLED` | `0` | Set to `1` to disable knowledge eviction. Eviction removes stale learnings that have decayed below the retention threshold. Disable if you want to keep all learnings indefinitely. |
 | `AGENTOPS_GITIGNORE_AUTO` | `1` | Set to `0` to prevent the session-start hook from auto-adding `.agents/` to `.gitignore`. Useful if you want to commit knowledge artifacts to your repo. |
diff --git a/docs/SKILLS.md b/docs/SKILLS.md
index 3ba5156c5..7b7010c99 100644
--- a/docs/SKILLS.md
+++ b/docs/SKILLS.md
@@ -446,7 +446,7 @@ These fire automatically and are not directly invoked:
 | `standards` | Language-specific coding standards (auto-loaded by /vibe, /implement) |
 | `shared` | Shared reference documents for multi-agent backends |
 | `beads` | Issue tracking reference |
-| `using-agentops` | Workflow guide (hook-capable auto-injection, explicit Codex startup fallback) |
+| `using-agentops` | Workflow guide (hook-capable silent state prep, explicit Codex startup fallback) |
 
 ---
 
diff --git a/docs/cli-skills-map.md b/docs/cli-skills-map.md
index 9317ce3b5..fbfc8734f 100644
--- a/docs/cli-skills-map.md
+++ b/docs/cli-skills-map.md
@@ -112,7 +112,7 @@ Which `ao` commands each hook invokes.
 
 | Hook File | Event | ao Commands |
 |-----------|-------|-------------|
-| **session-start.sh** | SessionStart | `extract`, `inject`, `lookup`, `notebook update`, `search` |
+| **session-start.sh** | SessionStart | `flywheel close-loop`, `knowledge brief`, `rpi cleanup` |
 | **session-end-maintenance.sh** | SessionEnd | `contradict`, `dedup`, `forge transcript`, `maturity`, `memory sync`, `notebook update`, `pool ingest` |
 | **ao-flywheel-close.sh** | Stop | `flywheel close-loop` |
 | **ratchet-advance.sh** | PostToolUse | `ratchet record` |
diff --git a/docs/contracts/hook-runtime-contract.md b/docs/contracts/hook-runtime-contract.md
index f4f82366c..9c9e5a444 100644
--- a/docs/contracts/hook-runtime-contract.md
+++ b/docs/contracts/hook-runtime-contract.md
@@ -17,7 +17,7 @@
 
 | Claude Event | Codex Equivalent | Mechanism |
 |-------------|-----------------|-----------|
-| `SessionStart` | `ao codex start` / `ao codex ensure-start` | Claude native hooks can recover a goal and surface a matched briefing automatically; Codex uses the explicit startup command path |
+| `SessionStart` | `ao codex start` / `ao codex ensure-start` | Claude native hooks can recover a goal and stage runtime state silently; Codex uses the explicit startup command path |
 | `SessionEnd` | `ao codex stop` / `ao codex ensure-stop` | Explicit command in skill epilogue or user invocation |
 | `Stop` | `ao codex stop --close-loop` | Explicit flywheel close via command |
 
@@ -25,7 +25,7 @@
 
 | Claude Event | Codex Adaptation | Notes |
 |-------------|-----------------|-------|
-| `UserPromptSubmit` | Skill preamble checks | Claude native hooks can capture first-prompt intake, emit prompt nudges, and run intent echo; Codex keeps equivalent guidance inside skill/runtime preambles instead of callbacks |
+| `UserPromptSubmit` | Skill preamble checks | Claude native hooks can capture first-prompt intake silently, emit prompt nudges, and run intent echo; Codex keeps equivalent guidance inside skill/runtime preambles instead of callbacks |
 | `TaskCompleted` | Skill epilogue validation | Task validation gates run inline after task completion |
 
 ### Unsupported (no Codex equivalent)
@@ -46,7 +46,7 @@ ao hook install  # Merges hooks.json into ~/.claude/settings.json
 - Full event manifest from `hooks/hooks.json`
 - `${CLAUDE_PLUGIN_ROOT}` path resolution
 - All events wired automatically
-- Native startup surface: `SessionStart` prefers matched goal-time briefings, and `UserPromptSubmit` can turn the first substantive prompt into factory intake when startup lacked a goal
+- `CLAUDE.md` remains the startup surface; hooks prepare factory state silently and keep only explicit safety nudges operator-facing
 
 ### Codex
 
diff --git a/docs/contracts/session-intelligence-trust-model.md b/docs/contracts/session-intelligence-trust-model.md
index e3b01a248..8d401e872 100644
--- a/docs/contracts/session-intelligence-trust-model.md
+++ b/docs/contracts/session-intelligence-trust-model.md
@@ -19,7 +19,7 @@ The goal is simple: do not let high-volume `.agents` capture turn into low-trust
 | finding | canonical | yes | yes | yes | yes | Promoted findings are the highest-trust reusable runtime signal. |
 | belief-book | canonical | yes | yes | yes | yes | Stable cross-domain doctrine promoted from healthy evidence. |
 | playbook | canonical | yes | yes | yes | yes | Generated operator workflows from healthy topics and promoted packets. |
-| knowledge-briefing | runtime-eligible | yes | no | no | yes | Preferred dynamic startup surface for a concrete goal; task-scoped, not universal policy. |
+| knowledge-briefing | runtime-eligible | yes | no | no | yes | Preferred explicit goal-time briefing surface for a concrete goal; task-scoped, not universal policy. |
 | planning-rule | canonical | yes | yes | yes | yes | Compiled planning rules are canonical prevention memory. |
 | known-risk | canonical | yes | yes | yes | yes | Compiled pre-mortem checks are canonical risk memory. |
 | next-work | runtime-eligible | yes | yes | yes | yes | Ranked next-work gives the next session continuity. |
@@ -44,7 +44,7 @@ Suppression does not mean deletion. These classes remain valid lookup, provenanc
 Session Intelligence should prefer this order when building startup or planning context:
 
 1. canonical artifacts
-2. matched knowledge briefings as the primary dynamic startup surface
+2. matched knowledge briefings when the operator or runtime explicitly assembles goal-time context
 3. runtime-eligible artifacts
 4. experimental families only after explicit health review
 5. discovery-only or archive-only artifacts only by explicit lookup or post-mortem use
diff --git a/docs/how-it-works.md b/docs/how-it-works.md
index 1be33fe00..102c4ac6c 100644
--- a/docs/how-it-works.md
+++ b/docs/how-it-works.md
@@ -108,7 +108,7 @@ The active runtime manifest currently declares **7 hook event sections** in `hoo
 
 | Hook surface | Trigger | What it does | Gap closed |
 |--------------|---------|--------------|------------|
-| Session start | `SessionStart` | Runs `session-start.sh` — startup maintenance, handoff recovery, matched briefing selection, and supporting retrieval | Durable learning (retrieval) |
+| Session start | `SessionStart` | Runs `session-start.sh` — startup maintenance, handoff recovery, and silent factory-state staging | Runtime continuity |
 | Session end maintenance | `SessionEnd` | Runs `session-end-maintenance.sh` (transcript mining, maturity management) and `athena-session-defrag.sh` (knowledge deduplication and defrag) | Durable learning (extraction), Loop closure |
 | Flywheel close | `Stop` | Runs `ao-flywheel-close.sh` — closes the feedback loop via `ao flywheel close-loop` | Loop closure |
 
@@ -116,7 +116,7 @@ The active runtime manifest currently declares **7 hook event sections** in `hoo
 
 | Hook surface | Trigger | What it does | Gap closed |
 |--------------|---------|--------------|------------|
-| Prompt guidance | `UserPromptSubmit` | Runs `factory-router.sh` (captures first-goal intake when startup had none), `prompt-nudge.sh` (ratchet nudges), and `intent-echo.sh` (confirms high-stakes intent) | Judgment validation |
+| Prompt guidance | `UserPromptSubmit` | Runs `factory-router.sh` (captures first-goal intake when startup had none), `prompt-nudge.sh` (ratchet nudges), and `intent-echo.sh` (confirms high-stakes intent) without adding startup briefings to the conversation | Judgment validation |
 | Pre-tool gates | `PreToolUse` | `pre-mortem-gate.sh` (blocks `/crank` without plan review), `commit-review-gate.sh` (pre-commit checks), `go-test-precommit.sh`, `git-worker-guard.sh` (worker isolation), `edit-knowledge-surface.sh`, `codex-parity-warn.sh` | Judgment validation |
 | Post-tool checks | `PostToolUse` | `write-time-quality.sh` (edit quality), `go-complexity-precommit.sh`, `go-vet-post-edit.sh`, `research-loop-detector.sh` (detects stalled loops), `context-monitor.sh` | Judgment validation, Loop closure |
 | Task completion gate | `TaskCompleted` | Runs `task-validation-gate.sh` — executes compiled constraints from `.agents/constraints/index.json` before accepting task completion | Judgment validation, Loop closure |
diff --git a/docs/workflows/session-lifecycle.md b/docs/workflows/session-lifecycle.md
index 7f0e48fe7..889d8f00c 100644
--- a/docs/workflows/session-lifecycle.md
+++ b/docs/workflows/session-lifecycle.md
@@ -14,8 +14,8 @@ Just describe what you want:
 
 | Say This | What Happens |
 |----------|--------------|
-| "Continue working on X" | `SessionStart` prefers a matched goal-time briefing, then loads supporting context and progress |
-| "I need to add Y" | If no startup goal exists yet, the first substantive prompt becomes factory intake and the runtime routes you toward `/rpi` |
+| "Continue working on X" | `CLAUDE.md` provides the startup surface; hooks recover handoff and stage any factory state silently |
+| "I need to add Y" | If no startup goal exists yet, the first substantive prompt can be captured as silent factory intake for later explicit `/rpi` or `ao knowledge brief` use |
 | "I'm done for today" | Saves progress, offers retrospective |
 | "What should I work on?" | Shows status, suggests next task |
 | "Where was I?" | Shows last session, current state, blockers |
@@ -62,7 +62,7 @@ with `ao codex ensure-stop`.
 
 | Mode | Start | Closeout | Notes |
 |------|-------|----------|-------|
-| Hook-capable | Natural language, `/session-start`, or startup hooks | Natural language, `/session-end`, or session-end hooks | Best fit for Claude/OpenCode when hooks are installed; startup prefers matched briefings and prompt-time intake can route into `/rpi` |
+| Hook-capable | Natural language, `/session-start`, or startup hooks | Natural language, `/session-end`, or session-end hooks | Best fit for Claude/OpenCode when hooks are installed; `CLAUDE.md` is the startup surface and hooks stage state silently |
 | Codex hookless fallback | `ao codex start` or skill-driven `ao codex ensure-start` | `ao codex stop` or skill-driven `ao codex ensure-stop` | No startup/session-end hook surface under `~/.codex`; lifecycle is explicit |
 | Manual fallback | `ao inject`, `ao lookup` | `ao forge transcript`, `ao flywheel close-loop` | Lowest-level portable path |
 
@@ -76,8 +76,8 @@ with `ao codex ensure-stop`.
 │  "Continue the API work" or /session-start                  │
 ├─────────────────────────────────────────────────────────────┤
 │  • Recover handoff / tracker goal when available            │
-│  • Prefer a matched goal-time briefing                      │
-│  • Add ranked learnings as supporting context               │
+│  • Keep operator framing in CLAUDE.md                       │
+│  • Stage factory goal / briefing files silently             │
 │  • Mark missing-goal sessions for prompt-time intake        │
 └─────────────────────────────────────────────────────────────┘
                             │
@@ -86,8 +86,8 @@ with `ao codex ensure-stop`.
 │                        WORK                                 │
 │  "Add the validation layer" or just start coding            │
 ├─────────────────────────────────────────────────────────────┤
-│  • First substantive prompt can become factory intake       │
-│  • Native hook path can build a briefing and route to /rpi  │
+│  • First substantive prompt can become silent factory intake│
+│  • Run /rpi or ao knowledge brief explicitly when needed    │
 │  • Implement features                                       │
 │  • Fix bugs                                                 │
 │  • Research when needed                                     │
diff --git a/hooks/factory-router.sh b/hooks/factory-router.sh
index 3407e378e..a3ee5b7ff 100755
--- a/hooks/factory-router.sh
+++ b/hooks/factory-router.sh
@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
-# factory-router.sh - UserPromptSubmit hook: convert first substantive prompt into
-# a goal-time briefing when SessionStart had no factory objective.
+# factory-router.sh - UserPromptSubmit hook: capture first substantive prompt as
+# factory state when SessionStart had no goal. This hook stays silent.
 
 [ "${AGENTOPS_HOOKS_DISABLED:-}" = "1" ] && exit 0
 [ "${AGENTOPS_FACTORY_ROUTER_DISABLED:-}" = "1" ] && exit 0
@@ -64,26 +64,6 @@ build_factory_briefing() {
     printf '%s' "$path"
 }
 
-read_factory_briefing() {
-    local path="$1"
-    local max_chars=2200
-    local content trimmed
-
-    [ -f "$path" ] || return 0
-    content=$(cat "$path" 2>/dev/null || true)
-    [ -n "$content" ] || return 0
-
-    if [ "${#content}" -gt "$max_chars" ]; then
-        trimmed="${content:0:$max_chars}"
-        content="${trimmed%
-*}
-
-*[briefing truncated by factory router]*"
-    fi
-
-    printf '%s' "$content"
-}
-
 extract_prompt() {
     local input="$1"
     if command -v jq >/dev/null 2>&1; then
@@ -157,37 +137,10 @@ printf '%s' "$GOAL" > "$GOAL_FILE" 2>/dev/null || true
 rm -f "$INTAKE_FLAG" 2>/dev/null || true
 printf '%s\n' "${CLAUDE_SESSION_ID:-unknown}" > "$ROUTER_FLAG" 2>/dev/null || true
 
-MESSAGE="Factory intake captured from the first substantive prompt.
-- **Goal:** ${GOAL}"
-
 if [ -n "$BRIEFING_PATH" ]; then
     printf '%s' "$BRIEFING_PATH" > "$BRIEFING_FILE" 2>/dev/null || true
-    BRIEFING_CONTENT="$(read_factory_briefing "$BRIEFING_PATH")"
-    MESSAGE="${MESSAGE}
-- **Primary briefing:** ${BRIEFING_PATH}
-- **Delivery lane:** Continue naturally, or run \`/rpi \"${GOAL}\"\`.
-
-Treat the briefing below as the primary dynamic surface for this session.
-
-<FACTORY_BRIEFING>
-${BRIEFING_CONTENT}
-</FACTORY_BRIEFING>"
 else
     rm -f "$BRIEFING_FILE" 2>/dev/null || true
-    MESSAGE="${MESSAGE}
-- **Primary briefing:** not available yet
-- **Delivery lane:** Run \`/rpi \"${GOAL}\"\` or continue with this objective.
-
-No matched knowledge briefing was available for the intake goal, so the captured goal above is the active factory objective."
-fi
-
-if command -v jq >/dev/null 2>&1; then
-    jq -n --arg msg "$MESSAGE" '{"hookSpecificOutput":{"hookEventName":"UserPromptSubmit","additionalContext":$msg}}'
-else
-    safe_msg=${MESSAGE//\\/\\\\}
-    safe_msg=${safe_msg//\"/\\\"}
-    safe_msg=$(printf '%s' "$safe_msg" | tr '\n' ' ')
-    echo "{\"hookSpecificOutput\":{\"hookEventName\":\"UserPromptSubmit\",\"additionalContext\":\"$safe_msg\"}}"
 fi
 
 exit 0
diff --git a/hooks/session-start.sh b/hooks/session-start.sh
index f47a97dec..5953a6888 100755
--- a/hooks/session-start.sh
+++ b/hooks/session-start.sh
@@ -1,8 +1,7 @@
 #!/usr/bin/env bash
 # AgentOps Session Start Hook
-# Creates .agents/ directories, consumes handoffs, and surfaces startup context.
-# In default factory mode, SessionStart prefers a matched knowledge briefing when
-# a goal is available; manual mode keeps a lighter startup surface.
+# Creates .agents/ directories, consumes handoffs, and prepares runtime state.
+# CLAUDE.md owns the operator-facing startup surface; this hook stays silent.
 
 # Kill switches
 [ "${AGENTOPS_HOOKS_DISABLED:-}" = "1" ] && exit 0
@@ -83,13 +82,6 @@ derive_lookup_query() {
     return 0
 }
 
-detect_lookup_bead() {
-    if ! command -v bd >/dev/null 2>&1; then
-        return 0
-    fi
-    run_with_timeout 1 bd current | head -1 | tr -d '\r' | sed 's/^ //; s/ $//'
-}
-
 build_factory_briefing() {
     local goal="$1"
     local output path
@@ -108,26 +100,6 @@ build_factory_briefing() {
     printf '%s' "$path"
 }
 
-read_factory_briefing() {
-    local path="$1"
-    local max_chars=2200
-    local content trimmed
-
-    [ -f "$path" ] || return 0
-    content=$(cat "$path" 2>/dev/null || true)
-    [ -n "$content" ] || return 0
-
-    if [ "${#content}" -gt "$max_chars" ]; then
-        trimmed="${content:0:$max_chars}"
-        content="${trimmed%
-*}
-
-*[briefing truncated by session-start hook]*"
-    fi
-
-    printf '%s' "$content"
-}
-
 write_environment_manifest() {
     local env_file="$AO_DIR/environment.json"
     local tmp_file git_branch head_sha git_dirty tools_json manifest_json
@@ -300,27 +272,11 @@ rpi/*
 EOF
 fi
 
-# Flywheel behavior
-INJECTED_KNOWLEDGE=""
 STARTUP_CONTEXT_MODE="$(resolve_startup_context_mode)"
 FACTORY_GOAL=""
 FACTORY_BRIEFING_PATH=""
-FACTORY_BRIEFING_CONTENT=""
-
-# Derive MEMORY.md path once (used for stale nudge)
-MEMORY_DIR="$HOME/.claude/projects"
-PROJECT_PATH=$(printf '%s' "$ROOT" | tr '/' '-')
-MEMORY_FILE="$MEMORY_DIR/$PROJECT_PATH/memory/MEMORY.md"
-MEMORY_AGE_DAYS=-1
-if [ -f "$MEMORY_FILE" ]; then
-    MTIME=$(stat -f %m "$MEMORY_FILE" 2>/dev/null || stat -c %Y "$MEMORY_FILE" 2>/dev/null || echo "")
-    if [ -n "$MTIME" ]; then
-        MEMORY_AGE_DAYS=$(( ($(date +%s) - MTIME) / 86400 ))
-    fi
-fi
 
 # Structured handoff consumption (ao handoff JSON artifacts)
-HANDOFF_CONTEXT=""
 if [ -d "$ROOT/.agents/handoff" ] && command -v jq &>/dev/null; then
     # Find newest unconsumed .json handoff (exclude .consumed.json and .consuming.json)
     HANDOFF_JSON=$(find "$ROOT/.agents/handoff" -maxdepth 1 -name 'handoff-*.json' \
@@ -340,52 +296,10 @@ if [ -d "$ROOT/.agents/handoff" ] && command -v jq &>/dev/null; then
                 "$CONSUMING" > "${CONSUMING}.tmp" 2>/dev/null \
                 && mv "${CONSUMING}.tmp" "${HANDOFF_JSON%.json}.consumed.json" 2>/dev/null \
                 && rm -f "$CONSUMING" 2>/dev/null
-            # Build injection context
-            HANDOFF_CONTEXT="### Handoff Context (${H_TYPE})"
-            [ -n "$H_GOAL" ] && HANDOFF_CONTEXT="${HANDOFF_CONTEXT}
-- **Goal:** ${H_GOAL}"
-            [ -n "$H_SUMMARY" ] && HANDOFF_CONTEXT="${HANDOFF_CONTEXT}
-- **Summary:** ${H_SUMMARY}"
-            [ -n "$H_CONTINUATION" ] && HANDOFF_CONTEXT="${HANDOFF_CONTEXT}
-- **Continue:** ${H_CONTINUATION}"
-            HANDOFF_CONTEXT="${HANDOFF_CONTEXT}
-- **Source:** ${HANDOFF_JSON}"
         fi
     fi
 fi
 
-# Predecessor handoff discovery
-PREDECESSOR_FILE="${GT_PREDECESSOR_HANDOFF:-}"
-if [ -z "$PREDECESSOR_FILE" ] && [ -d "$ROOT/.agents/handoff" ]; then
-    PREDECESSOR_FILE=$(ls -t "$ROOT/.agents/handoff/"*.md 2>/dev/null | head -1)
-fi
-
-# Build injection context (MEMORY.md is auto-loaded by Claude Code)
-INJECTED_KNOWLEDGE="MEMORY.md is auto-loaded by Claude Code for this project.
-For on-demand retrieval: \`ao search \"<query>\"\` or \`ao lookup --query \"<query>\"\`"
-if [ -n "$HANDOFF_CONTEXT" ]; then
-    INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-${HANDOFF_CONTEXT}"
-fi
-if [ -n "$PREDECESSOR_FILE" ] && [ -f "$PREDECESSOR_FILE" ]; then
-    INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-Predecessor handoff: ${PREDECESSOR_FILE}"
-fi
-
-# Point to knowledge signpost (agents search on demand)
-if [ -f "${ROOT:-.}/.agents/AGENTS.md" ]; then
-    INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-Knowledge artifacts are in \`.agents/\`. See \`.agents/AGENTS.md\` for navigation.
-Use \`ao lookup --query \"topic\"\` for learnings retrieval, or \`Grep\` in \`.agents/learnings/\` as fallback."
-else
-    INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-Knowledge artifacts are in \`.agents/\` (if populated).
-Use \`ao lookup --query \"topic\"\` for on-demand learnings retrieval."
-fi
-
 if [ "$STARTUP_CONTEXT_MODE" = "factory" ]; then
     FACTORY_GOAL="$(derive_lookup_query)"
     if [ -n "$FACTORY_GOAL" ]; then
@@ -393,138 +307,12 @@ if [ "$STARTUP_CONTEXT_MODE" = "factory" ]; then
         FACTORY_BRIEFING_PATH="$(build_factory_briefing "$FACTORY_GOAL")"
         if [ -n "$FACTORY_BRIEFING_PATH" ]; then
             printf '%s' "$FACTORY_BRIEFING_PATH" > "$ROOT/.agents/ao/factory-briefing.txt" 2>/dev/null || true
-            FACTORY_BRIEFING_CONTENT="$(read_factory_briefing "$FACTORY_BRIEFING_PATH")"
-            INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-### Factory Startup Surface
-- **Goal:** ${FACTORY_GOAL}
-- **Primary briefing:** ${FACTORY_BRIEFING_PATH}
-- **Delivery lane:** Continue naturally, or run \`/rpi \"${FACTORY_GOAL}\"\` for the full factory cycle.
-
-Treat the matched briefing below as the primary dynamic startup surface for this session. Use ranked learnings only as supporting evidence.
-
-<FACTORY_BRIEFING>
-${FACTORY_BRIEFING_CONTENT}
-</FACTORY_BRIEFING>"
         else
-            INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-### Factory Startup Surface
-- **Goal:** ${FACTORY_GOAL}
-- **Primary briefing:** not available yet
-- **Delivery lane:** Run \`/rpi \"${FACTORY_GOAL}\"\` or continue with this objective.
-
-No matched knowledge briefing was available at session start. Treat the goal above as the active factory objective; supporting learnings below are advisory rather than the primary brief."
+            rm -f "$ROOT/.agents/ao/factory-briefing.txt" 2>/dev/null || true
         fi
     else
         : > "$ROOT/.agents/ao/.factory-intake-needed" 2>/dev/null || true
-        INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-### Factory Startup Surface
-No startup goal was recovered from handoff or tracker state.
-Treat the first substantive user prompt as factory intake: compile a goal-time briefing, then route into \`/rpi\`."
     fi
 fi
 
-# Auto-retrieve and cite top learnings for this session (closes citation gap)
-if command -v ao &>/dev/null; then
-    LOOKUP_QUERY="$(derive_lookup_query)"
-    LOOKUP_BEAD="$(detect_lookup_bead)"
-    LOOKUP_CONTEXT=""
-    LOOKUP_ARGS=(lookup --limit 5)
-    if [ -n "$LOOKUP_QUERY" ]; then
-        LOOKUP_ARGS+=(--query "$LOOKUP_QUERY")
-        LOOKUP_CONTEXT="query=\"$LOOKUP_QUERY\""
-    fi
-    if [ -n "$LOOKUP_BEAD" ]; then
-        LOOKUP_ARGS+=(--bead "$LOOKUP_BEAD")
-        if [ -n "$LOOKUP_CONTEXT" ]; then
-            LOOKUP_CONTEXT="${LOOKUP_CONTEXT}, "
-        fi
-        LOOKUP_CONTEXT="${LOOKUP_CONTEXT}bead=${LOOKUP_BEAD}"
-    fi
-    if [ -z "$LOOKUP_CONTEXT" ]; then
-        LOOKUP_CONTEXT="task context unavailable"
-    fi
-    if [ -n "$LOOKUP_QUERY" ] || [ -n "$LOOKUP_BEAD" ]; then
-        FLYWHEEL_KNOWLEDGE=$(run_with_timeout 5 ao "${LOOKUP_ARGS[@]}" || true)
-    else
-        FLYWHEEL_KNOWLEDGE=""
-    fi
-    if [ -n "$FLYWHEEL_KNOWLEDGE" ]; then
-        INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-### Supporting Learnings (auto-retrieved: ${LOOKUP_CONTEXT})
-${FLYWHEEL_KNOWLEDGE}"
-    fi
-fi
-
-# Keep startup context lean: inject only fresh flywheel knowledge and a short skill pointer.
-SKILL_FILE="${PLUGIN_ROOT}/skills/using-agentops/SKILL.md"
-if [ -f "$SKILL_FILE" ]; then
-    using_agentops_hint="AgentOps workflow context is available. Use the Skill tool to read ${SKILL_FILE} when needed."
-else
-    using_agentops_hint="(AgentOps skill content unavailable at ${SKILL_FILE})"
-fi
-
-# Truncation (static cap — no notebook mode logic needed)
-MAX_INJECT_CHARS=4000
-if [ -n "$INJECTED_KNOWLEDGE" ] && [ "${#INJECTED_KNOWLEDGE}" -gt "$MAX_INJECT_CHARS" ]; then
-    # Truncate at last newline within budget (never mid-line)
-    trimmed="${INJECTED_KNOWLEDGE:0:$MAX_INJECT_CHARS}"
-    INJECTED_KNOWLEDGE="${trimmed%
-*}
-
-*[truncated by session-start hook]*"
-fi
-
-# Nudge agent if MEMORY.md is stale or missing
-if [ -n "${MEMORY_FILE:-}" ] && [ -f "$MEMORY_FILE" ] && [ "${MEMORY_AGE_DAYS:-0}" -gt 7 ]; then
-    INJECTED_KNOWLEDGE="${INJECTED_KNOWLEDGE}
-
-*Note: Your MEMORY.md hasn't been updated in ${MEMORY_AGE_DAYS} days. Consider running \`ao notebook update\` or updating it manually.*"
-fi
-
-if [ -n "$INJECTED_KNOWLEDGE" ]; then
-    full_content=$(cat <<HOOKCTX
-<AGENTOPS_CONTEXT>
-${INJECTED_KNOWLEDGE}
-
-${using_agentops_hint}
-</AGENTOPS_CONTEXT>
-HOOKCTX
-)
-else
-    full_content=$(cat <<HOOKCTX
-<AGENTOPS_CONTEXT>
-No prior flywheel knowledge was injected for this session.
-${using_agentops_hint}
-</AGENTOPS_CONTEXT>
-HOOKCTX
-)
-fi
-
-if command -v jq &>/dev/null; then
-    additional_context=$(printf '%s' "$full_content" | jq -Rs '.')
-    cat <<HOOKEOF
-{
-  "hookSpecificOutput": {
-    "hookEventName": "SessionStart",
-    "additionalContext": ${additional_context}
-  }
-}
-HOOKEOF
-else
-    # Minimal fallback — escape newlines and quotes
-    escaped=$(printf '%s' "$full_content" | sed 's/\\/\\\\/g; s/"/\\"/g' | tr '\n' ' ')
-    cat <<HOOKEOF
-{
-  "hookSpecificOutput": {
-    "hookEventName": "SessionStart",
-    "additionalContext": "${escaped}"
-  }
-}
-HOOKEOF
-fi
-
 exit 0
diff --git a/tests/hooks/hook-output-schema.bats b/tests/hooks/hook-output-schema.bats
index 9aed37293..522f20db2 100644
--- a/tests/hooks/hook-output-schema.bats
+++ b/tests/hooks/hook-output-schema.bats
@@ -55,20 +55,14 @@ assert_hook_schema() {
 }
 
 # ═══════════════════════════════════════════════════════════════════════
-# 1. session-start.sh (SessionStart) — registered
+# 1. session-start.sh (SessionStart) — registered but intentionally silent
 # ═══════════════════════════════════════════════════════════════════════
 
-@test "session-start: output conforms to hookSpecificOutput schema with hookEventName=SessionStart" {
+@test "session-start: produces no operator-facing JSON output" {
     run bash -c 'cd "$1" && CLAUDE_SESSION_ID="bats-schema-sess" bash "$2" 2>&1' \
         -- "$MOCK_REPO" "$HOOKS_DIR/session-start.sh"
     [ "$status" -eq 0 ]
-    [ -n "$output" ]
-    # session-start emits multi-line pretty-printed JSON (heredoc) possibly preceded
-    # by non-JSON diagnostic lines. Extract from first '{' to end and parse.
-    local json_blob
-    json_blob=$(echo "$output" | sed -n '/^{/,/^}/p' | jq -c '.' 2>/dev/null)
-    [ -n "$json_blob" ]
-    assert_hook_schema "$json_blob" "SessionStart"
+    [ -z "$output" ]
 }
 
 # ═══════════════════════════════════════════════════════════════════════
@@ -76,7 +70,7 @@ assert_hook_schema() {
 #    Only emits JSON when there is a nudge to give (ratchet gate active).
 # ═══════════════════════════════════════════════════════════════════════
 
-@test "factory-router: intake prompt output conforms to schema with hookEventName=UserPromptSubmit" {
+@test "factory-router: intake prompt stays silent while staging state" {
     local mock="$TMP_TEST_DIR/mock-factory-router"
     mkdir -p "$mock/.agents/ao" "$mock/bin"
     git -C "$mock" init -q >/dev/null 2>&1
@@ -99,8 +93,9 @@ EOF
     run bash -c 'cd "$1" && printf "%s" "$3" | PATH="$1/bin:$PATH" MOCK_FACTORY_BRIEFING_PATH="$2" AGENTOPS_STARTUP_CONTEXT_MODE=factory bash "$4" 2>&1' \
         -- "$mock" "$briefing_path" '{"prompt":"fix auth bootstrap"}' "$HOOKS_DIR/factory-router.sh"
     [ "$status" -eq 0 ]
-    [ -n "$output" ]
-    assert_hook_schema "$output" "UserPromptSubmit"
+    [ -z "$output" ]
+    [ "$(cat "$mock/.agents/ao/factory-goal.txt" 2>/dev/null)" = "fix auth bootstrap" ]
+    [ "$(cat "$mock/.agents/ao/factory-briefing.txt" 2>/dev/null)" = "$briefing_path" ]
 }
 
 @test "prompt-nudge: when nudge fires, output conforms to schema with hookEventName=UserPromptSubmit" {
@@ -363,11 +358,10 @@ AOEOF
     local hooks_json="$HOOKS_DIR/hooks.json"
     [ -f "$hooks_json" ]
 
-    # Parse hooks.json to build event-to-hook mapping for all known JSON emitters
+    # Parse hooks.json to build event-to-hook mapping for registered JSON emitters.
+    # SessionStart and factory-router are intentionally silent.
     local -A hook_event_map
     hook_event_map=(
-        ["session-start.sh"]="SessionStart"
-        ["factory-router.sh"]="UserPromptSubmit"
         ["prompt-nudge.sh"]="UserPromptSubmit"
         ["intent-echo.sh"]="UserPromptSubmit"
         ["commit-review-gate.sh"]="PreToolUse"
diff --git a/tests/hooks/test-hook-lifecycle.sh b/tests/hooks/test-hook-lifecycle.sh
index b3144dea1..4d58b3914 100755
--- a/tests/hooks/test-hook-lifecycle.sh
+++ b/tests/hooks/test-hook-lifecycle.sh
@@ -4,7 +4,7 @@
 # Verifies each hook produces correct side effects in the expected order.
 #
 # Lifecycle sequence:
-#   1. SessionStart   → session-start.sh (directory creation, JSON output)
+#   1. SessionStart   → session-start.sh (directory creation, silent runtime prep)
 #   2. PreToolUse(Read)  → citation-tracker.sh (citation recorded)
 #   3. PreToolUse(Bash)  → git-worker-guard.sh (worker git commit blocked)
 #   4. PostToolUse(Bash) → ratchet-advance.sh (advance suggestion on ratchet record)
@@ -163,21 +163,10 @@ else
     SESSION_JSON=$(echo "$SESSION_OUTPUT" | LC_ALL=C awk '/^[[:space:]]*\{/{found=1; buf=""} found{buf=buf $0 "\n"} /^[[:space:]]*\}/{if(found) last=buf; found=0} END{printf "%s", last}')
 fi
 
-if echo "$SESSION_JSON" | jq -e '.hookSpecificOutput.hookEventName == "SessionStart"' >/dev/null 2>&1; then
-    pass "step 1: emits SessionStart hookEventName"
+if [ -z "$SESSION_OUTPUT" ]; then
+    pass "step 1: session-start remains silent"
 else
-    # Fallback: grep raw output
-    if echo "$SESSION_OUTPUT" | grep -q '"SessionStart"'; then
-        pass "step 1: emits SessionStart hookEventName"
-    else
-        fail "step 1: emits SessionStart hookEventName"
-    fi
-fi
-
-if echo "$SESSION_JSON" | jq -e '.hookSpecificOutput.additionalContext' >/dev/null 2>&1; then
-    pass "step 1: emits additionalContext"
-else
-    fail "step 1: emits additionalContext"
+    fail "step 1: session-start remains silent"
 fi
 
 # Verify directory creation (session-start creates .agents/ dirs)
diff --git a/tests/hooks/test-hooks.bats b/tests/hooks/test-hooks.bats
index ad4ddcf3b..07c0d232b 100644
--- a/tests/hooks/test-hooks.bats
+++ b/tests/hooks/test-hooks.bats
@@ -28,15 +28,9 @@ teardown() {
 # session-start.sh / precompact-snapshot.sh
 # ═══════════════════════════════════════════════════════════════════════
 
-@test "session-start: emits SessionStart JSON" {
+@test "session-start: stays silent while preparing runtime state" {
     SESSION_RAW=$(bash "$HOOKS_DIR/session-start.sh" 2>/dev/null || true)
-    # Extract JSON from potentially noisy output
-    if echo "$SESSION_RAW" | grep -q '"hookEventName".*"SessionStart"'; then
-        true  # Found SessionStart in output
-    else
-        SESSION_JSON=$(echo "$SESSION_RAW" | LC_ALL=C awk '/^[[:space:]]*\{/{found=1; buf=""} found{buf=buf $0 "\n"} /^[[:space:]]*\}/{if(found) last=buf; found=0} END{printf "%s", last}')
-        echo "$SESSION_JSON" | jq -e '.hookSpecificOutput.hookEventName == "SessionStart"' >/dev/null 2>&1
-    fi
+    [ -z "$SESSION_RAW" ]
 }
 
 @test "session-start: kill switch suppresses output" {
@@ -52,7 +46,7 @@ teardown() {
     [ -d "$mock/.agents/research" ] && [ ! -d "$mock/subdir/.agents/research" ]
 }
 
-@test "session-start: factory mode builds and injects a matched briefing" {
+@test "session-start: factory mode stages a matched briefing without injecting it" {
     local mock="$TMP_TEST_DIR/mock-factory-start"
     mkdir -p "$mock/.agents/handoff" "$mock/bin"
     git -C "$mock" init -q >/dev/null 2>&1
@@ -94,8 +88,10 @@ EOF
     run bash -c 'cd "$1" && PATH="$1/bin:$PATH" AO_ARGS_FILE="$2" MOCK_FACTORY_BRIEFING_PATH="$3" AGENTOPS_STARTUP_CONTEXT_MODE=factory bash "$4" 2>&1' \
         -- "$mock" "$args_file" "$briefing_path" "$HOOKS_DIR/session-start.sh"
     [ "$status" -eq 0 ]
+    [ -z "$output" ]
     grep -q '^knowledge brief --json --goal stabilize auth startup$' "$args_file"
-    echo "$output" | jq -r '.hookSpecificOutput.additionalContext // ""' | grep -q '<FACTORY_BRIEFING>'
+    [ "$(cat "$mock/.agents/ao/factory-goal.txt" 2>/dev/null)" = "stabilize auth startup" ]
+    [ "$(cat "$mock/.agents/ao/factory-briefing.txt" 2>/dev/null)" = "$briefing_path" ]
 }
 
 @test "precompact-snapshot: emits additionalContext JSON" {
@@ -439,10 +435,12 @@ EOF
     run bash -c 'cd "$1" && printf "%s" "$4" | PATH="$1/bin:$PATH" AO_ARGS_FILE="$2" MOCK_FACTORY_BRIEFING_PATH="$3" AGENTOPS_STARTUP_CONTEXT_MODE=factory bash "$5" 2>&1' \
         -- "$mock" "$args_file" "$briefing_path" '{"prompt":"fix auth bootstrap"}' "$HOOKS_DIR/factory-router.sh"
     [ "$status" -eq 0 ]
+    [ -z "$output" ]
     grep -q '^knowledge brief --json --goal fix auth bootstrap$' "$args_file"
     [ ! -f "$mock/.agents/ao/.factory-intake-needed" ]
     [ "$(cat "$mock/.agents/ao/factory-goal.txt" 2>/dev/null)" = "fix auth bootstrap" ]
-    echo "$output" | jq -r '.hookSpecificOutput.additionalContext // ""' | grep -q '<FACTORY_BRIEFING>'
+    [ "$(cat "$mock/.agents/ao/factory-briefing.txt" 2>/dev/null)" = "$briefing_path" ]
+    [ -f "$mock/.agents/ao/.factory-router-fired" ]
 }
 
 # ═══════════════════════════════════════════════════════════════════════
diff --git a/tests/hooks/test-hooks.sh b/tests/hooks/test-hooks.sh
index d40124481..d627679a0 100755
--- a/tests/hooks/test-hooks.sh
+++ b/tests/hooks/test-hooks.sh
@@ -84,20 +84,12 @@ echo ""
 echo "=== session-start.sh / precompact-snapshot.sh ==="
 # ============================================================
 
-# Test 12: session-start emits valid JSON (extract last JSON object from output,
-# since ao extract may emit non-JSON to stdout before the hook's JSON)
+# Test 12: session-start stays silent while preparing runtime state
 SESSION_RAW=$(bash "$HOOKS_DIR/session-start.sh" 2>/dev/null || true)
-# Extract the last valid JSON block by finding the final { ... } spanning multiple lines
-SESSION_JSON=$(echo "$SESSION_RAW" | LC_ALL=C awk '/^[[:space:]]*\{/{found=1; buf=""} found{buf=buf $0 "\n"} /^[[:space:]]*\}/{if(found) last=buf; found=0} END{printf "%s", last}')
-if echo "$SESSION_JSON" | jq -e '.hookSpecificOutput.hookEventName == "SessionStart"' >/dev/null 2>&1; then
-    pass "session-start emits SessionStart JSON"
+if [ -z "$SESSION_RAW" ]; then
+    pass "session-start stays silent"
 else
-    # Fallback: check if hookEventName appears anywhere in raw output
-    if echo "$SESSION_RAW" | grep -q '"hookEventName".*"SessionStart"'; then
-        pass "session-start emits SessionStart JSON"
-    else
-        fail "session-start emits SessionStart JSON"
-    fi
+    fail "session-start stays silent"
 fi
 
 # Test 13: session-start kill switch suppresses output
@@ -115,7 +107,7 @@ else
     fail "session-start writes .agents to repo root"
 fi
 
-# Test 14b: session-start lookup uses handoff goal + active bead, not commit subject
+# Test 14b: session-start no longer performs hidden lookup injection
 MOCK_LOOKUP="$TMPDIR/mock-session-lookup"
 mkdir -p "$MOCK_LOOKUP/.agents/handoff" "$MOCK_LOOKUP/bin"
 git -C "$MOCK_LOOKUP" init -q >/dev/null 2>&1
@@ -130,8 +122,8 @@ cat > "$MOCK_LOOKUP/.agents/handoff/handoff-20260322T160000Z.json" <<'EOF'
   "id": "handoff-test",
   "created_at": "2026-03-22T16:00:00Z",
   "type": "manual",
-  "goal": "task-scoped lookup queries",
-  "summary": "use the handoff goal for retrieval"
+  "goal": "task-scoped silent startup",
+  "summary": "startup should materialize state without injected lookup chatter"
 }
 EOF
 cat > "$MOCK_LOOKUP/bin/ao" <<'EOF'
@@ -139,39 +131,23 @@ cat > "$MOCK_LOOKUP/bin/ao" <<'EOF'
 if [ -n "${AO_ARGS_FILE:-}" ]; then
     printf '%s\n' "$*" >> "$AO_ARGS_FILE"
 fi
-if [ "${1:-}" = "lookup" ]; then
-    printf '[lookup] stub result\n'
-fi
 exit 0
 EOF
-cat > "$MOCK_LOOKUP/bin/bd" <<'EOF'
-#!/usr/bin/env bash
-if [ "${1:-}" = "current" ]; then
-    printf 'ag-73u.5\n'
-fi
-exit 0
-EOF
-chmod +x "$MOCK_LOOKUP/bin/ao" "$MOCK_LOOKUP/bin/bd"
+chmod +x "$MOCK_LOOKUP/bin/ao"
 AO_ARGS_FILE="$MOCK_LOOKUP/ao-args.log"
 LOOKUP_OUTPUT=$(cd "$MOCK_LOOKUP" && PATH="$MOCK_LOOKUP/bin:$PATH" AO_ARGS_FILE="$AO_ARGS_FILE" bash "$HOOKS_DIR/session-start.sh" 2>/dev/null || true)
-if grep -q '^lookup --limit 5 --query task-scoped lookup queries --bead ag-73u.5$' "$AO_ARGS_FILE"; then
-    pass "session-start lookup is scoped by handoff goal and bead"
+if ! grep -qE '^(lookup|knowledge brief)\b' "$AO_ARGS_FILE" 2>/dev/null; then
+    pass "session-start performs no hidden lookup or briefing work in manual mode"
 else
-    fail "session-start lookup is scoped by handoff goal and bead"
+    fail "session-start performs no hidden lookup or briefing work in manual mode"
 fi
-if ! grep -q 'commit subject should not drive lookup' "$AO_ARGS_FILE"; then
-    pass "session-start no longer falls back to commit subject when task context exists"
+if [ -z "$LOOKUP_OUTPUT" ]; then
+    pass "session-start remains silent in manual mode"
 else
-    fail "session-start no longer falls back to commit subject when task context exists"
-fi
-LOOKUP_CONTEXT=$(echo "$LOOKUP_OUTPUT" | jq -r '.hookSpecificOutput.additionalContext // ""' 2>/dev/null)
-if echo "$LOOKUP_CONTEXT" | grep -q 'auto-retrieved: query="task-scoped lookup queries", bead=ag-73u.5'; then
-    pass "session-start reports lookup scope in injected context"
-else
-    fail "session-start reports lookup scope in injected context"
+    fail "session-start remains silent in manual mode"
 fi
 
-# Test 14c: session-start prefers a factory briefing when startup goal exists
+# Test 14c: session-start can stage a factory briefing silently when startup goal exists
 MOCK_FACTORY_START="$TMPDIR/mock-factory-start"
 mkdir -p "$MOCK_FACTORY_START/.agents/handoff" "$MOCK_FACTORY_START/bin"
 git -C "$MOCK_FACTORY_START" init -q >/dev/null 2>&1
@@ -221,11 +197,16 @@ if grep -q '^knowledge brief --json --goal stabilize auth startup$' "$AO_ARGS_FI
 else
     fail "session-start builds a factory briefing from startup goal"
 fi
-FACTORY_CONTEXT=$(echo "$FACTORY_OUTPUT" | jq -r '.hookSpecificOutput.additionalContext // ""' 2>/dev/null)
-if echo "$FACTORY_CONTEXT" | grep -q '<FACTORY_BRIEFING>' && echo "$FACTORY_CONTEXT" | grep -q 'stabilize auth startup'; then
-    pass "session-start injects the factory briefing as primary startup surface"
+if [ "$(cat "$MOCK_FACTORY_START/.agents/ao/factory-goal.txt" 2>/dev/null)" = "stabilize auth startup" ] && \
+   [ "$(cat "$MOCK_FACTORY_START/.agents/ao/factory-briefing.txt" 2>/dev/null)" = "$MOCK_FACTORY_BRIEFING_PATH" ]; then
+    pass "session-start records staged factory state"
 else
-    fail "session-start injects the factory briefing as primary startup surface"
+    fail "session-start records staged factory state"
+fi
+if [ -z "$FACTORY_OUTPUT" ]; then
+    pass "session-start does not inject startup context in factory mode"
+else
+    fail "session-start does not inject startup context in factory mode"
 fi
 
 # Test 15: precompact emits JSON when data exists
@@ -1819,11 +1800,16 @@ if [ ! -f "$FACTORY_ROUTER_MOCK/.agents/ao/.factory-intake-needed" ] && \
 else
     fail "factory-router records the captured goal and clears intake state"
 fi
-ROUTER_CONTEXT=$(echo "$ROUTER_OUTPUT" | jq -r '.hookSpecificOutput.additionalContext // ""' 2>/dev/null)
-if echo "$ROUTER_CONTEXT" | grep -q '<FACTORY_BRIEFING>' && echo "$ROUTER_CONTEXT" | grep -q '/rpi "fix auth bootstrap"'; then
-    pass "factory-router emits factory routing guidance with briefing"
+if [ "$(cat "$FACTORY_ROUTER_MOCK/.agents/ao/factory-briefing.txt" 2>/dev/null)" = "$MOCK_FACTORY_BRIEFING_PATH" ] && \
+   [ -f "$FACTORY_ROUTER_MOCK/.agents/ao/.factory-router-fired" ]; then
+    pass "factory-router stages briefing state and dedup marker"
 else
-    fail "factory-router emits factory routing guidance with briefing"
+    fail "factory-router stages briefing state and dedup marker"
+fi
+if [ -z "$ROUTER_OUTPUT" ]; then
+    pass "factory-router stays silent"
+else
+    fail "factory-router stays silent"
 fi
 if jq -e '.hooks.UserPromptSubmit[].hooks[] | select(.command | contains("factory-router.sh"))' "$HOOKS_DIR/hooks.json" >/dev/null 2>&1; then
     pass "factory-router.sh wired in hooks.json"
diff --git a/tests/integration/test-hook-chain.sh b/tests/integration/test-hook-chain.sh
index 0bde2f61e..912f107f8 100755
--- a/tests/integration/test-hook-chain.sh
+++ b/tests/integration/test-hook-chain.sh
@@ -142,42 +142,21 @@ else
     fail "session-start.sh exits with code 0 (got $EXIT_CODE)"
 fi
 
-if [ -n "$OUTPUT" ]; then
-    pass "session-start.sh produces output"
+if [ -z "$OUTPUT" ]; then
+    pass "session-start.sh stays silent"
 else
-    fail "session-start.sh produces output"
+    fail "session-start.sh stays silent"
 fi
 
 # ============================================================
 echo ""
-echo "=== Test 2: session-start.sh JSON output validation ==="
+echo "=== Test 2: session-start.sh silent runtime prep ==="
 # ============================================================
 
-# Validate JSON structure
-if echo "$OUTPUT" | jq . >/dev/null 2>&1; then
-    pass "session-start.sh produces valid JSON"
+if [ -z "$OUTPUT" ]; then
+    pass "session-start.sh emits no operator-facing context"
 else
-    fail "session-start.sh produces valid JSON"
-fi
-
-# Check for required fields
-if echo "$OUTPUT" | jq -e '.hookSpecificOutput.hookEventName' >/dev/null 2>&1; then
-    pass "JSON contains hookEventName field"
-else
-    fail "JSON contains hookEventName field"
-fi
-
-if echo "$OUTPUT" | jq -e '.hookSpecificOutput.additionalContext' >/dev/null 2>&1; then
-    pass "JSON contains additionalContext field"
-else
-    fail "JSON contains additionalContext field"
-fi
-
-EVENT_NAME=$(echo "$OUTPUT" | jq -r '.hookSpecificOutput.hookEventName')
-if [ "$EVENT_NAME" = "SessionStart" ]; then
-    pass "hookEventName is 'SessionStart'"
-else
-    fail "hookEventName is 'SessionStart' (got: $EVENT_NAME)"
+    fail "session-start.sh emits no operator-facing context"
 fi
 
 # ============================================================