diff --git a/CHANGELOG.md b/CHANGELOG.md
index 1ae1e81..442813b 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -5,6 +5,30 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [1.5.0] - 2026-03-20
+
+### Added
+- **Stop-hook haiku summarization** — AI context summaries are now derived from
+  Claude's actual response (`last_assistant_message`) via the Stop hook, delivering
+  ~100% of the time (up from 0.5% with the previous inline strategy). Summaries
+  reflect what Claude *did*, not just what was asked.
+- **Two-phase context** — first-5-words appears immediately on prompt submission;
+  a refined haiku summary overwrites after the turn completes.
+- **Git commit message capture** — commit subject line automatically written to
+  context file on `💾 Committed`, providing a human-written task summary at zero cost.
+- **Context sanitization** — shell prompt prefixes (`(venv) user@host %`) are
+  stripped from first-5-words when users paste terminal output.
+
+### Fixed
+- Clear `CLAUDECODE` env var in haiku subprocess to prevent nested claude call
+  failures.
+
+### Removed
+- **Inline AI context strategy** — `--ai-context-strategy` flag,
+  `--append-system-prompt` injection, `CCP_TASK_SUMMARY` markers, and
+  `inline_captured` sentinel files are all removed. `--ai-context` is now a
+  simple boolean toggle.
+
 ## [1.2.0] - 2026-03-05
 
 ### Added
@@ -98,6 +122,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - **SECURITY.md** — responsible disclosure policy.
 - **Issue templates** (bug report + feature request) and PR template.
 
+[1.5.0]: https://github.com/brianruggieri/claude-code-pulse/compare/v1.2.0...v1.5.0
 [1.2.0]: https://github.com/brianruggieri/claude-code-pulse/compare/v1.1.0...v1.2.0
 [1.1.0]: https://github.com/brianruggieri/claude-code-pulse/compare/v1.0.0...v1.1.0
 [1.0.0]: https://github.com/brianruggieri/claude-code-pulse/releases/tag/v1.0.0
diff --git a/README.md b/README.md
index 4d2f0cc..40deca7 100644
--- a/README.md
+++ b/README.md
@@ -41,7 +41,9 @@ Each pane title is independent — project, branch, task, and live status update
 - **Priority-based display**. Errors always show first. Completions (tests passed, committed) immediately override whatever else is showing.
 - **Hook-based**. Uses Claude Code's own hook events for status. No output parsing, no regex, just structured JSON.
 - **Session tracking**. Save and re-open sessions by title with `--goto`.
-- **AI task summaries**. `--ai-context` refines your prompt into a clean 3-5 word label shown in the title. Two strategies: `haiku` (separate claude-haiku call, fully invisible) or `inline` (zero extra API calls, uses the already-outgoing session). Both are opt-in.
+- **AI task summaries**. `--ai-context` refines each turn into a clean 3-5 word label shown in the title. Uses claude-haiku to summarize after each turn completes. Opt-in only.
+- **Git commit context**. When Claude commits, the commit subject line automatically appears as the task context — no flag needed.
+- **Context sanitization**. Shell prompt prefixes (`$`, `%`, `user@host`, etc.) are automatically stripped from pasted terminal content before building the title.
 - **Tested on iTerm2, Terminal.app, and tmux**. Detection logic for WezTerm, Ghostty, and Kitty is included but unverified.
 - **Full Claude passthrough**. Every Claude Code flag works exactly as expected.
 
@@ -75,14 +77,11 @@ ccp --permission-mode bypassPermissions
 # AI-summarized task labels in the title (opt-in, uses your subscription)
 ccp --ai-context "Fix the auth bug"
 
-# Inline strategy — zero extra API calls (one echo visible per session)
-ccp --ai-context --ai-context-strategy inline "Fix the auth bug"
-
 # Debug mode — structured JSONL logging + live Pulse Monitor (iTerm2: auto-opens in split pane)
 ccp --debug-ccp "Fix the auth bug"
 ```
 
-> See [AI Task Summaries](#ai-task-summaries) for the `--ai-context` flag and strategy options.
+> See [AI Task Summaries](#ai-task-summaries) for the `--ai-context` flag.
 
 ## Installation
 
@@ -232,30 +231,19 @@ Precedence: `--status-profile` flag > `CCP_STATUS_PROFILE` env var > `quiet` def
 
 ## AI Task Summaries
 
-With `--ai-context`, ccp distills your prompt into a clean 3-5 word label shown in the title. Without it, the first words of your prompt show as-is.
-
-Two strategies are available:
+With `--ai-context`, ccp generates a clean 3-5 word label for each conversation turn. Without it, the first words of your prompt show as-is.
 
-| | `haiku` (default) | `inline` |
-|---|---|---|
-| Extra API calls | 1 per prompt | 0 |
-| Visible to user | No | One `echo` per session |
-| Summary context | Prompt text only | Full codebase + history |
-| Reliability | High | Model-dependent |
+How it works: when Claude finishes responding, the Stop hook sends a truncated excerpt of the response to claude-haiku for a 3-5 word summary. The summary overwrites the first-5-words placeholder in the title.
 
 ```bash
-# Haiku strategy (default) — invisible, uses your subscription
+# Enable AI summaries (opt-in, uses your subscription)
 ccp --ai-context "PR #89 - Fix auth"
 
-# Inline strategy — zero extra API calls, one echo visible per session
-ccp --ai-context --ai-context-strategy inline "PR #89 - Fix auth"
-
-# Always on, via environment variables
+# Always on, via environment variable
 export CCP_ENABLE_AI_CONTEXT=true
-export CCP_AI_CONTEXT_STRATEGY=inline   # optional, default is haiku
 ```
 
-See [docs/ai-context.md](docs/ai-context.md) for the full trade-off analysis.
+See [docs/ai-context.md](docs/ai-context.md) for the full details.
 
 ## Claude Flag Passthrough
 
@@ -284,7 +272,7 @@ Any unrecognized `-*` flag is also forwarded, so new Claude flags just work.
 ### Full forwarded flag list
 
 **No value:**
-`-c` / `--continue`, `--dangerously-skip-permissions`, `--allow-dangerously-skip-permissions`, `--verbose`, `--ide`, `--no-ide`, `--fork-session`, `--chrome`, `--no-chrome`, `--no-session-persistence`, `--include-partial-messages`, `--replay-user-messages`, `--strict-mcp-config`, `--mcp-debug`
+`-c` / `--continue`, `-p` / `--print`, `--dangerously-skip-permissions`, `--allow-dangerously-skip-permissions`, `--verbose`, `--ide`, `--no-ide`, `--fork-session`, `--chrome`, `--no-chrome`, `--no-session-persistence`, `--include-partial-messages`, `--replay-user-messages`, `--strict-mcp-config`, `--mcp-debug`
 
 **Required value:**
 `--model`, `--permission-mode`, `--effort`, `--system-prompt`, `--append-system-prompt`, `--debug-file`, `--session-id`, `--output-format`, `--input-format`, `--settings`, `--setting-sources`, `--fallback-model`, `--max-budget-usd`, `--json-schema`, `--agent`
@@ -430,8 +418,7 @@ my-app (bug/login-crash)     | login bug  | ⬆️ Pushing
 | `--auto-title` | Auto-detect title from git branch (this is the default) |
 | `--no-dynamic` | Static title only, no live updates |
 | `--status-profile quiet\|verbose` | Which events to surface (default: `quiet`) |
-| `--ai-context` | Summarize prompts into a 3-5 word title label (opt-in) |
-| `--ai-context-strategy haiku\|inline` | Summary strategy: `haiku` (default, invisible) or `inline` (zero extra API calls) |
+| `--ai-context` | Summarize each turn into a 3-5 word title label (opt-in, uses your subscription) |
 | `--debug-ccp` | Structured JSONL debug logging to `~/.config/claude-code-pulse/logs/`; auto-opens Pulse Monitor in iTerm2 |
 | `--goto TITLE` | Re-open a previous session by title |
 | `--list`, `-l` | List active ccp sessions |
@@ -445,8 +432,7 @@ my-app (bug/login-crash)     | login bug  | ⬆️ Pushing
 | Variable | Default | Description |
 |----------|---------|-------------|
 | `CCP_STATUS_PROFILE` | `quiet` | Default status profile |
-| `CCP_ENABLE_AI_CONTEXT` | `false` | Always-on AI prompt summarization |
-| `CCP_AI_CONTEXT_STRATEGY` | `haiku` | Summary strategy: `haiku` or `inline` |
+| `CCP_ENABLE_AI_CONTEXT` | `false` | Always-on AI context summarization |
 | `CCP_DEBUG` | `false` | Always-on structured JSONL debug logging |
 
 ## Documentation
diff --git a/bin/ccp b/bin/ccp
index 460e2a5..0dfbb1e 100755
--- a/bin/ccp
+++ b/bin/ccp
@@ -60,16 +60,11 @@ ${BLUE}ccp Options:${NC}
     --auto-title            Auto-detect title from git branch (default)
     --no-dynamic            Disable dynamic updates (static title)
     --status-profile MODE   Status surface: quiet (default) or verbose
-    --ai-context            Summarize prompts in title via claude-haiku (opt-in,
-                            uses your Claude subscription — see docs/ai-context.md)
+    --ai-context            Summarize tasks in title via claude-haiku (opt-in,
+                            uses your Claude subscription)
                             Privacy: prompt text is sent to Anthropic to generate
                             the summary. Do not enable if prompts contain secrets.
                             Also: export CCP_ENABLE_AI_CONTEXT=true for always-on
-    --ai-context-strategy MODE
-                            How AI context summaries are generated:
-                              haiku   — separate claude-haiku call (default)
-                              inline  — system prompt injection, zero extra API calls
-                            Also: export CCP_AI_CONTEXT_STRATEGY=inline for always-on
     --debug-ccp             Enable structured JSONL debug logging to
                             ~/.config/claude-code-pulse/logs/
                             In iTerm2, auto-opens Pulse Monitor in a split pane.
@@ -166,7 +161,6 @@ main() {
     local enable_dynamic=true
     local status_profile_cli=""
     local enable_ai_context="${CCP_ENABLE_AI_CONTEXT:-false}"
-    local ai_context_strategy="${CCP_AI_CONTEXT_STRATEGY:-haiku}"
     local enable_debug_ccp="${CCP_DEBUG:-false}"
     local claude_args=()
     local restored_title=""
@@ -205,15 +199,6 @@ main() {
                 enable_ai_context=true
                 shift
                 ;;
-            --ai-context-strategy)
-                shift
-                if [[ $# -lt 1 ]]; then
-                    log_error "Usage: ${SCRIPT_NAME} --ai-context-strategy haiku|inline"
-                    exit 1
-                fi
-                ai_context_strategy="$1"
-                shift
-                ;;
             --status-profile)
                 shift
                 if [[ $# -lt 1 ]]; then
@@ -397,18 +382,6 @@ main() {
     esac
     export CCP_STATUS_PROFILE="${status_profile}"
 
-    # Resolve AI context strategy with precedence:
-    #   CLI flag > environment variable > default haiku.
-    case "${ai_context_strategy}" in
-        haiku|inline)
-            ;;
-        *)
-            log_error "Invalid AI context strategy '${ai_context_strategy}'. Use: haiku or inline."
-            exit 1
-            ;;
-    esac
-    export CCP_AI_CONTEXT_STRATEGY="${ai_context_strategy}"
-
     # Setup debug logging if enabled
     if [[ "${enable_debug_ccp}" = true ]]; then
         local log_dir="${STATE_DIR}/logs"
@@ -439,8 +412,7 @@ main() {
             --arg dir "${directory}" \
             --arg profile "${status_profile}" \
             --arg ai_ctx "${enable_ai_context}" \
-            --arg ai_strat "${ai_context_strategy}" \
-            '{ts:$ts, src:"session", pid:$pid, event:"session_start", title:$title, dir:$dir, profile:$profile, ai_context:$ai_ctx, ai_strategy:$ai_strat}' \
+            '{ts:$ts, src:"session", pid:$pid, event:"session_start", title:$title, dir:$dir, profile:$profile, ai_context:$ai_ctx}' \
             >> "${CCP_DEBUG_LOG}" 2>/dev/null || true
     fi
 
@@ -466,7 +438,7 @@ main() {
         cleanup_monitor
         teardown_ccp_hooks "${ccp_settings_file}"
         rm -f "${CCP_STATUS_FILE:-}" "${CCP_CONTEXT_FILE:-}" "${CCP_BRANCH_FILE:-}" \
-             "${CCP_AGENTS_FILE:-}" "${STATE_DIR}/inline_captured.$$"
+             "${CCP_AGENTS_FILE:-}"
     }
     trap '_ccp_cleanup' EXIT
 
@@ -484,33 +456,14 @@ main() {
         log_info "Dynamic titles: ${GREEN}enabled${NC}"
         log_info "Status profile: ${CCP_STATUS_PROFILE}"
         if [[ "${enable_ai_context}" = true ]]; then
-            if [[ "${ai_context_strategy}" == "inline" ]]; then
-                log_info "AI context:     ${GREEN}enabled${NC} (inline — summary via system prompt, no extra API calls)"
-            else
-                log_info "AI context:     ${GREEN}enabled${NC} (prompts summarized via claude-haiku — uses your subscription)"
-            fi
+            log_info "AI context:     ${GREEN}enabled${NC} (task summaries via claude-haiku — uses your subscription)"
         fi
         if [[ "${enable_debug_ccp}" = true ]]; then
             log_info "Debug log:      ${YELLOW}${CCP_DEBUG_LOG}${NC}"
         fi
     fi
     if [[ ${#claude_args[@]} -gt 0 ]]; then
-        local display_args=()
-        local skip_next=false
-        for arg in "${claude_args[@]}"; do
-            if [[ "${skip_next}" == true ]]; then
-                skip_next=false
-                continue
-            fi
-            if [[ "${arg}" == "--append-system-prompt" ]]; then
-                skip_next=true
-                continue
-            fi
-            display_args+=("${arg}")
-        done
-        if [[ ${#display_args[@]} -gt 0 ]]; then
-            log_info "Claude args:    ${display_args[*]}"
-        fi
+        log_info "Claude args:    ${claude_args[*]}"
     fi
     echo ""
 
@@ -554,21 +507,12 @@ main() {
             | awk '{print $1}') || _welcome_name=""
         [[ -z "${_welcome_name}" ]] && _welcome_name="${USER:-}"
         if [[ -n "${_welcome_name}" ]]; then
-            printf '%s' "👋 Welcome back, ${_welcome_name}" > "${CCP_STATUS_FILE}"
+            printf '%s' "👋 Welcome back, ${_welcome_name}!" > "${CCP_STATUS_FILE}"
         fi
 
         # Inject hooks into the project's .claude/settings.local.json
         ccp_settings_file=$(setup_ccp_hooks "${directory}" "${CCP_HOOK_RUNNER}")
 
-        # Inline AI context: inject a system prompt instruction asking Claude
-        # to echo a structured task summary as its first action.  The summary
-        # is captured by the PostToolUse hook (see hook_runner.sh post-tool).
-        if [[ "${enable_ai_context}" = true && "${ai_context_strategy}" == "inline" ]]; then
-            local _ccp_inline_prompt
-            _ccp_inline_prompt="[CCP Terminal Title] As your very first action, run a single bash command: echo 'CCP_TASK_SUMMARY_$$:' followed by a 3-5 word title-case summary of the user's task. Example: echo 'CCP_TASK_SUMMARY_$$:Fix JWT Validation Bug'. Do this once, then proceed with your work normally."
-            claude_args+=("--append-system-prompt" "${_ccp_inline_prompt}")
-        fi
-
         # Start background title updater, then run Claude directly.
         title_updater "${title}"
 
diff --git a/bin/ccp-watch b/bin/ccp-watch
index aceeec6..6e1e277 100755
--- a/bin/ccp-watch
+++ b/bin/ccp-watch
@@ -72,9 +72,7 @@ fi
 last_status=""
 last_status_ts=""
 last_event_ts=""
-pending_summaries=0
 delivered_summaries=0
-inline_miss_warned=false
 total_errors=0
 total_events=0
 
@@ -284,26 +282,11 @@ tail -f "${log}" | while IFS= read -r line; do
         fi
     fi
 
-    # ── Inline summary tracking ────────────────────────────────────────
+    # ── AI summary tracking ────────────────────────────────────────────
     case "${event}" in
-        ai_context_pending)
-            pending_summaries=$((pending_summaries + 1))
-            ;;
         ai_summary)
             delivered_summaries=$((delivered_summaries + 1))
-            alert "OK" "Summary delivered: '${summary}' (${delivered_summaries}/${pending_summaries})"
-            ;;
-        inline_marker_miss)
-            # Only warn once per session — inline strategy is best-effort,
-            # Claude may skip the summary echo entirely.
-            if [[ "${inline_miss_warned}" = false && ${pending_summaries} -gt ${delivered_summaries} ]]; then
-                miss_count=$(jq -n "[inputs | select(.event == \"inline_marker_miss\")] | length" "${log}" 2>/dev/null) || miss_count=0
-                if [[ ${miss_count} -ge 20 ]]; then
-                    cmd=$(printf '%s' "${line}" | jq -r '.command // ""' 2>/dev/null | head -c 40) || cmd=""
-                    alert "INFO" "Inline summary not delivered after ${miss_count} Bash calls (best-effort, using first-5-words fallback)"
-                    inline_miss_warned=true
-                fi
-            fi
+            alert "OK" "Summary delivered: '${summary}' (${delivered_summaries} total)"
             ;;
     esac
 
@@ -339,11 +322,9 @@ tail -f "${log}" | while IFS= read -r line; do
         echo -e "  ${MAGENTA}${BOLD}Session Summary${NC}"
         echo -e "  Events:     ${total_events}"
         echo -e "  Errors:     ${total_errors}"
-        echo -e "  Summaries:  ${delivered_summaries}/${pending_summaries} delivered"
+        echo -e "  Summaries:  ${delivered_summaries} delivered"
         if [[ ${total_errors} -gt 0 ]]; then
             echo -e "  ${RED}${BOLD}Issues detected — review log with: ccp-watch --list${NC}"
-        elif [[ ${pending_summaries} -gt ${delivered_summaries} ]]; then
-            echo -e "  ${YELLOW}Dropped summaries detected${NC}"
         else
             echo -e "  ${GREEN}Clean session.${NC}"
         fi
diff --git a/docs/ai-context.md b/docs/ai-context.md
index 2f809ad..811cfd1 100644
--- a/docs/ai-context.md
+++ b/docs/ai-context.md
@@ -4,7 +4,7 @@
 
 ## What It Does
 
-When enabled, ccp watches the first message you send to Claude in each session and generates a concise 3–5 word label for it. That label appears in the terminal title alongside the current status:
+When enabled, ccp generates a concise 3–5 word label for each conversation turn and shows it in the terminal title:
 
 ```
 ✳ my-project (main) | Fix Auth Bug | ✏️ Editing
@@ -26,120 +26,80 @@ With `--ai-context`, it's refined into a clean label:
 
 ## How It Works
 
-There are two strategies for generating the AI summary:
+Context updates happen in two phases:
 
-### Haiku Strategy (default)
+### Phase 1: First-5-words (immediate)
 
-When you send a prompt, `hook_runner.sh`'s UserPromptSubmit handler:
+When you send a prompt, `hook_runner.sh`'s UserPromptSubmit handler writes the first 5 words of your prompt to the context file immediately. The title updates within one polling cycle.
 
-1. Writes the first 5 words of your prompt to the context file immediately (instant title update)
-2. Fires a **detached background subprocess** with `claude --print --model claude-haiku-4-5-20251001`
-3. Sends this exact prompt to Haiku:
+### Phase 2: Haiku summary (after turn completes)
 
-```
-Summarize this developer task in 3-5 words. Title-case. No punctuation. No quotes. Reply with only the words, nothing else. Task: <your prompt text>
-```
+When Claude finishes responding, the Stop hook fires with `last_assistant_message` — the full text of Claude's response. A detached background subprocess:
 
-4. When Haiku responds (~1–3 seconds), overwrites the context file with the summary
+1. Truncates the response to first 500 + last 500 characters
+2. Sends it to `claude --print --model claude-haiku-4-5-20251001` with a summarization prompt
+3. Haiku returns a 3–5 word title-case summary
+4. The summary overwrites the first-5-words placeholder in the context file
 5. The title_updater picks it up on the next heartbeat
 
 The subprocess is fully detached (`& disown`). It never blocks Claude, never delays your session, and never writes to the terminal.
 
-### Inline Strategy (`--ai-context-strategy inline`)
-
-Instead of a separate API call, ccp piggybacks on the main Claude session:
+## Additional Context Sources
 
-1. Writes the first 5 words of your prompt to the context file immediately (instant title update)
-2. Injects a lightweight `--append-system-prompt` instruction into Claude's launch args
-3. Claude reads the instruction and, as its very first action, runs: `echo 'CCP_TASK_SUMMARY:Fix JWT Auth Bug'`
-4. The PostToolUse hook captures the echo output, extracts the summary after the `CCP_TASK_SUMMARY:` marker
-5. Writes the summary to the context file
-6. The title_updater picks it up on the next heartbeat
+Two features provide context automatically, independent of `--ai-context`:
 
-**Key differences from Haiku:**
-- **Zero extra API calls** — the summary is part of the main Claude response
-- **Zero extra cost** — no additional subscription usage
-- **Better summaries** — Claude has full codebase context, not just the prompt text
-- **One visible echo** — the `echo 'CCP_TASK_SUMMARY:...'` appears in Claude's tool use output (minimal, one-time)
-- **Model-dependent** — Claude may occasionally skip the instruction (falls back to first-5-words)
+### Git commit message capture
 
-## What Data Is Sent
+When Claude runs `git commit` and the PostToolUse hook detects a successful commit, the commit subject line is extracted and written to the context file. This replaces the prompt-based context with a human-readable summary — e.g., `feat: add OAuth support` appears alongside `💾 Committed`.
 
-### Haiku Strategy
+### Context sanitization
 
-**Exactly your prompt text** — the full message you typed into Claude Code, after stripping the project name if it appears in the title.
+Shell prompt prefixes are automatically stripped from pasted terminal content before extracting the first-5-words placeholder. Patterns like `user@host`, `(venv)`, `$`, `%`, `#`, and `>` at the start of the prompt are removed so prefix tokens don't waste the 5-word budget.
 
-No other data is sent. No file contents, no tool outputs, no conversation history. Just the user-facing message you already typed.
+## What Data Is Sent
 
-### Inline Strategy
+**Claude's response text** — the `last_assistant_message` field from the Stop hook, truncated to first 500 + last 500 characters.
 
-**Nothing extra.** Your prompt is already being sent to Claude as part of the normal session. The only addition is a ~250-character system prompt instruction appended via `--append-system-prompt`. No additional user data leaves your machine.
+No file contents beyond what Claude already processed. No conversation history. Just a truncated excerpt of what Claude just said.
 
 ## Subscription Usage
 
-### Haiku Strategy
-
-This strategy calls `claude --print` once per user message. That's one Haiku API call per prompt you send.
+This feature calls `claude --print` once per conversation turn (when Claude stops responding). That's one Haiku API call per turn.
 
 - **Model:** `claude-haiku-4-5-20251001` (the most cost-effective model)
 - **Call type:** Non-interactive, single-turn (`--print` flag)
-- **Context sent:** Your prompt text only (one message, no history)
+- **Context sent:** Truncated response excerpt (max ~1000 chars)
 - **Response:** 3–5 words (minimal token usage)
 
-If you use Claude via subscription (Claude.ai Pro/Max), this counts against your subscription usage. If you use API keys, this counts against your API bill. The per-call cost is minimal (Haiku is Anthropic's cheapest model), but it fires every time you send a message, so usage scales with how much you type.
+If you use Claude via subscription (Claude.ai Pro/Max), this counts against your subscription usage. If you use API keys, this counts against your API bill. The per-call cost is minimal (Haiku is Anthropic's cheapest model), but it fires every turn, so usage scales with conversation length.
 
 **This is exactly why the feature is opt-in.** We don't think it's appropriate to silently consume your subscription on your behalf. You should choose to enable it knowing what it does.
 
-### Inline Strategy
-
-**Zero additional subscription usage.** The summary is generated as part of the main Claude session you're already paying for. The only overhead is a single `echo` command in Claude's first response — negligible additional tokens.
-
 ## Privacy
 
-### Haiku Strategy
-
 The data flow is:
 
 ```
-Your prompt text → claude CLI (your auth) → Anthropic API → 3-5 word summary
+Claude's response text (truncated) → claude CLI (your auth) → Anthropic API → 3-5 word summary
 ```
 
-- Your prompt is processed under your own Claude account and credentials
+- Claude's response is processed under your own Claude account and credentials
 - The summary request is subject to Anthropic's standard privacy policy — the same policy that applies to every message you send to Claude
 - ccp does not see, log, or store the data — the call happens entirely through your local `claude` CLI binary
 
 Nothing is routed through ccp's infrastructure (there isn't any). The subprocess calls the same `claude` binary you use interactively.
 
-### Inline Strategy
-
-The data flow is:
-
-```
-Your prompt text (already going) + ~250-char system instruction → Claude session → echo summary
-```
-
-- Your prompt is sent to Anthropic **once** (the main session), not twice
-- The only addition is a short system prompt instruction appended via `--append-system-prompt`
-- No data leaves your machine beyond what already would without ccp
-- ccp does not see, log, or store the data — the echo output is captured by the PostToolUse hook and written to a local temp file
+**Note:** Claude's response text is sent to Haiku for summarization. If the response contains sensitive information (secrets, credentials, personal data), that content will be included in the summarization request.
 
 ## How to Enable
 
 ```bash
-# Flag (per session) — uses Haiku strategy by default
+# Flag (per session)
 ccp --ai-context "My task"
 
-# Inline strategy (no extra API calls — recommended)
-ccp --ai-context --ai-context-strategy inline "My task"
-
 # Environment variable (always-on — add to ~/.zshrc or ~/.bashrc)
 export CCP_ENABLE_AI_CONTEXT=true
 ccp "My task"  # AI context active without --ai-context flag
-
-# Inline strategy always-on
-export CCP_ENABLE_AI_CONTEXT=true
-export CCP_AI_CONTEXT_STRATEGY=inline
-ccp "My task"
 ```
 
 The flag and env var are equivalent. The env var is the recommended approach if you always want AI context active.
@@ -147,22 +107,16 @@ The flag and env var are equivalent. The env var is the recommended approach if
 When enabled, ccp prints a disclosure at startup:
 
 ```
-# Haiku strategy
 Dynamic titles: enabled
 Status profile: quiet
 AI context:     enabled (prompts summarized via claude-haiku — uses your subscription)
-
-# Inline strategy
-Dynamic titles: enabled
-Status profile: quiet
-AI context:     enabled (inline — summary via system prompt, no extra API calls)
 ```
 
 ## How to Disable
 
 Don't pass `--ai-context`. That's all.
 
-If you had previously set `CCP_ENABLE_AI_CONTEXT=true` in your shell profile, remove that line. If you also set `CCP_AI_CONTEXT_STRATEGY=inline`, remove that too.
+If you had previously set `CCP_ENABLE_AI_CONTEXT=true` in your shell profile, remove that line.
 
 ## Without AI Context
 
diff --git a/docs/dynamic-titles.md b/docs/dynamic-titles.md
index d647f88..bf679e3 100644
--- a/docs/dynamic-titles.md
+++ b/docs/dynamic-titles.md
@@ -112,7 +112,7 @@ Fired when a Bash tool finishes. We inspect the output for completion keywords:
 |-----------|--------|
 | `N tests passed` / `N specs passed` | `✅ Tests passed` |
 | `N tests failed` / `N specs failing` | `❌ Tests failed` |
-| git output starts with `[` (commit hash) | `💾 Committed` |
+| git output starts with `[` (commit hash) | `💾 Committed` (+ commit subject written to context file) |
 
 **Completion events always win**, regardless of current priority. A `✅ Tests passed` (priority 60) immediately overrides any active status like `🧪 Testing` (priority 80).
 
@@ -128,14 +128,13 @@ Fired when a tool fails:
 ### UserPromptSubmit — User sends a message
 
 1. Clears the status file (removes any stale status such as the startup welcome message)
-2. Writes the first 5 words of the user's prompt to the context file as a placeholder
-3. If `--ai-context` is enabled: spawns a background call to `claude --print` with model `claude-haiku-4-5-20251001`
-4. Haiku distills the full prompt into a 3–5 word summary (opt-in only — uses your subscription)
-5. Summary overwrites the context file when ready (~1–3 seconds)
+2. Sanitizes the prompt (strips shell prompt prefixes like `$`, `%`, `user@host` from pasted content)
+3. Writes the first 5 words of the sanitized prompt to the context file as an immediate placeholder
 
 ### Stop — Claude finishes responding
 
-Clears the status file. The monitor shows `💤 Idle`.
+1. Clears the status file. The monitor shows `💤 Idle`.
+2. If `--ai-context` is enabled: reads `last_assistant_message` from the Stop hook payload, truncates to first 500 + last 500 characters, and spawns a detached background call to `claude --print --model claude-haiku-4-5-20251001` for a 3–5 word summary. The summary overwrites the context file when ready (~1–3 seconds).
 
 ### Event Hooks (Verbose Mode)
 
@@ -150,7 +149,7 @@ Additional hooks fire for lifecycle and permission events. These are only surfac
 | SessionStart | `🚀 Starting` |
 | PreCompact | `📦 Compacting` |
 | SubagentStart | `🤖 Delegating` |
-| SubagentStop | `✅ Complete` |
+| SubagentStop | `✅ Subagent finished` |
 | TeammateIdle | `⏸️ Teammate idle` |
 | ConfigChange | `⚙️ Configuring` |
 
@@ -172,7 +171,8 @@ my-project (main) | Fix Auth Bug | ✏️ Editing
 - Always preserved
 
 **Task summary**
-- First 5 words of your prompt (always); refined to 3–5 word semantic summary if `--ai-context` is enabled
+- First 5 words of your prompt (immediate); refined to 3–5 word summary after the turn completes if `--ai-context` is enabled
+- On `💾 Committed`, the commit subject line replaces the prompt-based context automatically
 - Only appears after the first UserPromptSubmit
 - Before that, the "Welcome" status is shown instead
 
@@ -242,9 +242,9 @@ Claude Code then reads the updated settings file and registers the hooks.
 
 As Claude Code runs:
 - PreToolUse fires → hook_runner sets PreToolUse status
-- PostToolUse fires → hook_runner sets PostToolUse/completion status
-- UserPromptSubmit fires → hook_runner sets Thinking; spawns distillation only if `--ai-context` enabled
-- Stop fires → hook_runner clears status (monitor shows Idle)
+- PostToolUse fires → hook_runner sets PostToolUse/completion status; captures commit subject on `💾 Committed`
+- UserPromptSubmit fires → hook_runner sets Thinking; writes first-5-words context
+- Stop fires → hook_runner clears status (monitor shows Idle); spawns haiku summarization if `--ai-context` enabled
 
 ### Teardown
 
@@ -284,7 +284,7 @@ A separate subshell keeps this work isolated and allows the main session to retu
 
 ### Why Haiku distillation?
 
-User prompts can be long and wandering. When `--ai-context` is enabled, after each user message we run `claude --print` (non-interactive, one turn) with the full prompt to generate a 3–5 word summary. This summary fits neatly in the terminal title and gives context at a glance ("Fix Auth Bug", "Refactor Logging", etc.).
+User prompts can be long and wandering. When `--ai-context` is enabled, after Claude finishes responding the Stop hook sends a truncated excerpt of the response to `claude --print` (non-interactive, one turn) for a 3–5 word summary. This summary fits neatly in the terminal title and gives context at a glance ("Fix Auth Bug", "Refactor Logging", etc.).
 
 This feature is **opt-in** and explicitly uses your Claude subscription. See [docs/ai-context.md](ai-context.md) for the full details, privacy implications, and how to enable it.
 
diff --git a/docs/inline-context-spec.md b/docs/inline-context-spec.md
deleted file mode 100644
index 33ddd14..0000000
--- a/docs/inline-context-spec.md
+++ /dev/null
@@ -1,154 +0,0 @@
-# Inline AI Context: Specification
-
-> **Status:** Implemented  
-> **Feature flag:** `--ai-context-strategy inline` / `CCP_AI_CONTEXT_STRATEGY=inline`
-
-## Problem
-
-The existing `--ai-context` feature sends a **separate** `claude --print --model claude-haiku` API call to summarize each user prompt. This works, but has trade-offs:
-
-- **Extra cost** — one Haiku call per prompt (subscription or API credits).
-- **Extra latency** — 1–3 seconds for the background call to return.
-- **Reduced context** — Haiku only sees the raw prompt text, not the codebase or conversation history.
-- **Duplicate data** — the prompt is sent to Anthropic twice (once for the main session, once for summarization).
-
-## Solution: Inline Context Strategy
-
-Piggyback on the **already-outgoing** Claude request. Instead of a separate API call, CCP injects a lightweight system prompt instruction into the main Claude session via `--append-system-prompt`. Claude generates a task summary as part of its normal first-turn processing, and CCP captures it through the existing PostToolUse hook.
-
-### How It Works
-
-```
-┌──────────────────────────────────────────────────────────────────┐
-│  User types prompt → goes to Claude (as always)                  │
-│                                                                  │
-│  CCP adds --append-system-prompt with a one-line instruction:    │
-│  "echo 'CCP_TASK_SUMMARY:Brief Title' as your first action"     │
-│                                                                  │
-│  Claude reads user prompt + full codebase context                │
-│  Claude echoes: CCP_TASK_SUMMARY:Fix JWT Auth Bug                │
-│  Claude proceeds with normal work                                │
-│                                                                  │
-│  PostToolUse hook sees the echo output                           │
-│  hook_runner.sh extracts the summary after the marker            │
-│  Writes to CCP_CONTEXT_FILE                                     │
-│  Monitor picks it up → title updates                             │
-└──────────────────────────────────────────────────────────────────┘
-```
-
-### Data Flow
-
-```
-User prompt ──→ Claude (with appended system prompt)
-                    │
-                    ├─→ echo 'CCP_TASK_SUMMARY:Fix JWT Auth'  (first action)
-                    │       │
-                    │       └─→ PostToolUse hook → hook_runner.sh
-                    │               │
-                    │               └─→ extract summary → CCP_CONTEXT_FILE
-                    │                       │
-                    │                       └─→ monitor → terminal title
-                    │
-                    └─→ normal Claude work continues
-```
-
-## System Prompt Instruction
-
-The injected instruction is deliberately minimal to avoid wasting context window space:
-
-```
-[CCP Terminal Title] As your very first action, run: echo 'CCP_TASK_SUMMARY:' followed by
-a 3-5 word title-case summary of the user's task. Example: echo 'CCP_TASK_SUMMARY:Fix JWT
-Validation Bug'. Do this once, then proceed normally.
-```
-
-**Key design choices:**
-
-1. **Structured marker** (`CCP_TASK_SUMMARY:`) — unambiguous prefix for reliable parsing. Not a plausible substring of any real command output.
-2. **Bash echo** — the simplest possible tool use. No file writes, no permissions, no side effects. The echo goes to stdout and is captured by the PostToolUse hook's `tool_response` field.
-3. **One-shot** — "Do this once" prevents repeated summaries on follow-up prompts.
-4. **Minimal instruction** — under 250 characters. Negligible context window impact.
-
-## Comparison: Haiku vs. Inline
-
-| Aspect | Haiku (`haiku`) | Inline (`inline`) |
-|---|---|---|
-| **API calls** | +1 Haiku call per prompt | Zero extra calls |
-| **Cost** | Haiku tokens (small but nonzero) | Zero additional cost |
-| **Latency** | 1–3s background call | ~0s (part of main response) |
-| **Summary quality** | Haiku-quality (good) | Main-model quality (better — full context) |
-| **Context available** | Prompt text only | Full codebase + conversation history |
-| **User visibility** | Invisible (detached subprocess) | One brief `echo` in Claude's output |
-| **Privacy** | Prompt sent twice to Anthropic | Prompt sent once (already going) |
-| **Reliability** | High (dedicated call) | Model-dependent (Claude may skip it) |
-| **Fallback** | First 5 words if Haiku fails | First 5 words if echo is skipped |
-
-## Configuration
-
-```bash
-# Enable inline strategy (per session)
-ccp --ai-context --ai-context-strategy inline "My task"
-
-# Enable inline strategy (always-on via env vars)
-export CCP_ENABLE_AI_CONTEXT=true
-export CCP_AI_CONTEXT_STRATEGY=inline
-ccp "My task"
-```
-
-The strategy setting only matters when `--ai-context` is enabled. Without `--ai-context`, both strategies are inactive and the title shows the raw first-5-words fallback.
-
-## Privacy & Security
-
-### What changes
-
-- **Haiku strategy**: prompt text is sent to Anthropic twice (main session + Haiku call).
-- **Inline strategy**: prompt text is sent to Anthropic once (main session only). A ~250-character system prompt instruction is appended.
-
-### What doesn't change
-
-- No data is sent to any third party (all calls go through your local `claude` CLI).
-- No data is logged, stored, or transmitted by CCP.
-- The summary stays local (written to a temp file in `~/.config/claude-code-pulse/`).
-- The feature is opt-in (requires `--ai-context`).
-
-### Transparency
-
-When the inline strategy is active, CCP prints at startup:
-
-```
-AI context:     enabled (inline — summary via system prompt, no extra API calls)
-```
-
-The system prompt instruction is visible in Claude's conversation context (Claude Code shows system prompt contents when asked).
-
-## Terms & Convention Compliance
-
-1. **`--append-system-prompt`** is Claude Code's official mechanism for extending the system prompt. CCP uses it as documented.
-2. **PostToolUse hooks** are Claude Code's official mechanism for observing tool outputs. CCP reads, never modifies.
-3. **The appended instruction does not override or conflict** with the user's own `--system-prompt` or `--append-system-prompt` flags. Claude Code concatenates multiple `--append-system-prompt` values.
-4. **Claude's behavior is not silently altered** — the system prompt is part of the visible configuration, and the echo output appears in Claude's tool use history.
-
-## Failure Modes
-
-| Scenario | Behavior |
-|---|---|
-| Claude ignores the instruction | First-5-words fallback remains in the title (same as no AI context) |
-| Claude includes extra text in the echo | Marker prefix parsing extracts only the summary portion |
-| Claude echoes on every prompt | Only the latest summary is used (atomic overwrite) |
-| User also passes `--append-system-prompt` | Both are applied (Claude Code concatenates) |
-| Inline + Haiku both enabled | Inline takes precedence (Haiku subprocess is skipped) |
-
-## Implementation Summary
-
-### `bin/ccp`
-- Parses `--ai-context-strategy` flag and `CCP_AI_CONTEXT_STRATEGY` env var
-- When inline: injects `--append-system-prompt` with the summary instruction into `claude_args`
-- Exports `CCP_AI_CONTEXT_STRATEGY` for `hook_runner.sh`
-
-### `lib/hook_runner.sh`
-- `post-tool` handler: detects `CCP_TASK_SUMMARY:` in Bash tool_response, extracts summary, writes to `CCP_CONTEXT_FILE`
-- `user-prompt` handler: skips Haiku subprocess when strategy is `inline` (summary will come from the main session)
-
-### `docs/ai-context.md`
-- Documents both strategies with clear comparison
-- Explains when to use each one
diff --git a/docs/usage.md b/docs/usage.md
index 60fa324..3a0d9df 100644
--- a/docs/usage.md
+++ b/docs/usage.md
@@ -124,8 +124,7 @@ Any unrecognized `-*` flag is also forwarded, so you can pass new Claude flags w
 | `--auto-title` | Detect title from git branch (default behavior) |
 | `--no-dynamic` | Static title only — disable monitoring |
 | `--status-profile quiet\|verbose` | Status surface (default: quiet) |
-| `--ai-context` | Summarize prompts into a 3-5 word title label (opt-in) |
-| `--ai-context-strategy haiku\|inline` | Summary strategy: `haiku` (default, invisible) or `inline` (zero extra API calls) |
+| `--ai-context` | Summarize each turn into a 3-5 word title label (opt-in, uses your subscription) |
 | `--goto TITLE` | Re-open previous session by title search |
 | `--list`, `-l` | List all active ccp sessions |
 | `--help`, `-h` | Show help |
diff --git a/lib/hook_runner.sh b/lib/hook_runner.sh
index 30a83ac..18bb2d8 100644
--- a/lib/hook_runner.sh
+++ b/lib/hook_runner.sh
@@ -8,6 +8,7 @@
 #        CCP_CONTEXT_FILE - path to write task context (user prompt)
 #        CCP_STATUS_PROFILE - quiet (default) or verbose
 #        CCP_DEBUG_LOG    - optional path for debug output
+#        CCP_CLAUDE_BIN   - override claude binary path (for testing)
 #
 # Always exits 0 — must never block or fail Claude Code.
 
@@ -117,7 +118,7 @@ _status_priority() {
         *"Awaiting approval"*) echo 88 ;;
         *"Input needed"*)  echo 85 ;;
         *Building*|*Testing*|*Installing*) echo 80 ;;
-        *Pushing*|*Pulling*|*Merging*) echo 75 ;;
+        *Pushing*|*Pulling*|*Merging*|*Rebasing*|*Cherry-picking*) echo 75 ;;
         *Docker*|*Thinking*|*Delegating*) echo 70 ;;
         *Editing*)         echo 65 ;;
         *"Tests passed"*|*Committed*|*Completed*|*"Subagent finished"*) echo 60 ;;
@@ -419,87 +420,109 @@ case "${mode}" in
             exit 0
         fi
 
+        # Strip shell prompt prefixes from pasted terminal content before
+        # extracting first-5-words so prefix tokens don't fill the word budget.
+        sanitized_prompt=""
+        sanitized_prompt=$(printf '%s' "${raw_prompt}" \
+            | sed 's/^([^)]*)[[:space:]]*//' \
+            | sed 's/^[a-zA-Z0-9._-]*@[a-zA-Z0-9._-]*[[:space:]]*//' \
+            | sed 's/^[a-zA-Z0-9._-]*[[:space:]]*[%$#>][[:space:]]*//' \
+            | sed 's/^[%$#>][[:space:]]*//')
+
         # Write first-5-words placeholder immediately so the title updates at once
         initial=""
-        initial=$(printf '%s' "${raw_prompt}" \
+        initial=$(printf '%s' "${sanitized_prompt}" \
             | awk '{n=(NF<5?NF:5); for(i=1;i<=n;i++) printf "%s%s",$i,(i<n?" ":""); print ""}')
+
         [[ -n "${initial}" ]] && atomic_write "${CCP_CONTEXT_FILE}" "${initial}"
         _dbg_event "context_set" "context=${initial}" "source=first-5-words"
+        ;;
 
-        # AI context summarization is opt-in (--ai-context flag / CCP_ENABLE_AI_CONTEXT=true).
-        # It sends your prompt text to claude-haiku and counts against your subscription.
-        # Skip unless explicitly enabled.
-        if [[ "${CCP_ENABLE_AI_CONTEXT:-false}" != "true" ]]; then
-            _dbg "AI context summarization not enabled (use --ai-context to enable)"
-            exit 0
+    stop)
+        [[ -z "${CCP_STATUS_FILE:-}" ]] && exit 0
+        _dbg_event "status_cleared" "trigger=stop"
+        _dbg "clearing status file"
+        # Empty status signals idle to the monitor on the next heartbeat
+        _cur=""
+        if [[ -f "${CCP_STATUS_FILE}" ]]; then
+            _cur=$(cat "${CCP_STATUS_FILE}" 2>/dev/null) || _cur=""
+        fi
+        if [[ -z "${_cur}" ]]; then
+            _dbg_event "status_dedup" "status="
+        else
+            atomic_write "${CCP_STATUS_FILE}" ""
         fi
 
-        # Inline strategy: the summary is produced by the main Claude session
-        # (via --append-system-prompt injection) and captured in the post-tool
-        # handler.  No separate API call needed — skip the Haiku subprocess.
-        if [[ "${CCP_AI_CONTEXT_STRATEGY:-haiku}" == "inline" ]]; then
-            _dbg_event "ai_context_pending" "strategy=inline" "waiting_for=post-tool CCP_TASK_SUMMARY marker"
-            _dbg "AI context strategy=inline — skipping haiku subprocess"
+        # AI context summarization — summarize what Claude just did.
+        # Uses last_assistant_message from the Stop hook payload (added in
+        # Claude Code v1.0.17).  The detached subprocess survives the hook's
+        # 5-second timeout — it runs independently after disown.
+        if [[ "${CCP_ENABLE_AI_CONTEXT:-false}" != "true" ]] || \
+           [[ -z "${CCP_CONTEXT_FILE:-}" ]]; then
             exit 0
         fi
 
-        # Background AI distillation — rewrites the context file with a proper
-        # 3-5 word semantic summary once the haiku call completes (~1-3s).
-        # CCP vars are unset inside the subshell so any hooks fired by the child
-        # claude process become no-ops (hook_runner exits early when vars unset).
+        last_msg=""
+        last_msg=$(printf '%s' "${json_input}" | jq -r '.last_assistant_message // ""' 2>/dev/null) || true
+        [[ -z "${last_msg}" ]] && exit 0
+
+        _dbg_event "ai_context_stop" "msg_len=${#last_msg}"
+
+        # Truncate: first 500 + last 500 chars
+        _truncated="${last_msg}"
+        if [[ ${#last_msg} -gt 1000 ]]; then
+            _first="${last_msg:0:500}"
+            _last="${last_msg: -500}"
+            _truncated="${_first} ... ${_last}"
+        fi
+
         _ccp_ctx="${CCP_CONTEXT_FILE}"
-        _ccp_proj="${CCP_PROJECT_NAME:-}"
-        _ccp_raw="${raw_prompt}"
+        _ccp_truncated="${_truncated}"
+        _ccp_claude_bin="${CCP_CLAUDE_BIN:-}"
+        # Snapshot current context so the subprocess can detect if a newer
+        # prompt has arrived before it finishes (race guard).
+        _ccp_expected_ctx=""
+        if [[ -f "${CCP_CONTEXT_FILE}" ]]; then
+            _ccp_expected_ctx=$(< "${CCP_CONTEXT_FILE}") || _ccp_expected_ctx=""
+        fi
         (
             set +e
-            unset CCP_STATUS_FILE CCP_CONTEXT_FILE
-
-            claude_bin=$(command -v claude 2>/dev/null \
-                || command -v claude-code 2>/dev/null || echo "")
-            [[ -z "${claude_bin}" ]] && exit 0
-
-            # Strip project name from prompt so the summary doesn't repeat it
-            task_text="${_ccp_raw}"
-            if [[ -n "${_ccp_proj}" ]]; then
-                task_text=$(printf '%s' "${task_text}" \
-                    | sed "s/ (${_ccp_proj})[^,]*,\{0,1\}[[:space:]]*/  /g" \
-                    | sed "s/${_ccp_proj}//g" \
-                    | sed 's/^[[:space:]]*//' | sed 's/[[:space:]]*$//')
+            # CLAUDECODE: env var set by Claude Code itself. Must be cleared or
+            # nested `claude --print` calls fail (discovered by Quickchat.ai).
+            # CCP_*: unset so hooks fired by the child claude process become no-ops.
+            unset CLAUDECODE CCP_ENABLE_AI_CONTEXT CCP_STATUS_FILE CCP_CONTEXT_FILE
+
+            claude_bin="${_ccp_claude_bin}"
+            if [[ -z "${claude_bin}" ]]; then
+                claude_bin=$(command -v claude 2>/dev/null \
+                    || command -v claude-code 2>/dev/null || echo "")
             fi
+            [[ -z "${claude_bin}" ]] && exit 0
 
-            summary=$(printf 'Summarize this developer task in 3-5 words. Title-case. No punctuation. No quotes. Reply with only the words, nothing else. Task: %s' \
-                    "${task_text}" \
+            summary=$(printf 'Summarize what this AI assistant just did in 3-5 words. Title-case. No punctuation. No quotes. Reply with only the words, nothing else.\n\nResponse:\n%s' \
+                    "${_ccp_truncated}" \
                 | "${claude_bin}" --print --model claude-haiku-4-5-20251001 \
                 2>/dev/null | head -1 \
                 | sed 's/^[[:space:]]*//' | sed 's/[[:space:]]*$//') || true
 
-            [[ -n "${summary}" ]] && atomic_write "${_ccp_ctx}" "${summary}"
-            _dbg_event "ai_summary" "summary=${summary}" "strategy=haiku"
-            _dbg "distilled: ${summary}"
+            # Race guard: if a newer prompt has updated the context file
+            # since we started, skip the write to avoid overwriting it.
+            if [[ -n "${summary}" ]]; then
+                _current_ctx=""
+                [[ -f "${_ccp_ctx}" ]] && _current_ctx=$(< "${_ccp_ctx}") || _current_ctx=""
+                if [[ "${_current_ctx}" == "${_ccp_expected_ctx}" ]]; then
+                    atomic_write "${_ccp_ctx}" "${summary}"
+                fi
+            fi
+            _dbg_event "ai_summary" "summary=${summary}" "strategy=stop-haiku"
         ) &
         disown 2>/dev/null || true
         ;;
 
-    stop)
-        [[ -z "${CCP_STATUS_FILE:-}" ]] && exit 0
-        _dbg_event "status_cleared" "trigger=stop"
-        _dbg "clearing status file"
-        # Empty status signals idle to the monitor on the next heartbeat
-        _cur=""
-        if [[ -f "${CCP_STATUS_FILE}" ]]; then
-            _cur=$(cat "${CCP_STATUS_FILE}" 2>/dev/null) || _cur=""
-        fi
-        if [[ -z "${_cur}" ]]; then
-            _dbg_event "status_dedup" "status="
-        else
-            atomic_write "${CCP_STATUS_FILE}" ""
-        fi
-        ;;
-
     post-tool)
         # All three output files are checked: status detection writes to
-        # CCP_STATUS_FILE, inline AI context writes to CCP_CONTEXT_FILE, and
-        # branch refresh writes to CCP_BRANCH_FILE.  Skip only if none are set.
+        # CCP_STATUS_FILE, commit message capture writes to CCP_CONTEXT_FILE,
+        # and branch refresh writes to CCP_BRANCH_FILE.  Skip only if none are set.
         [[ -z "${CCP_STATUS_FILE:-}" && -z "${CCP_CONTEXT_FILE:-}" && -z "${CCP_BRANCH_FILE:-}" ]] && exit 0
 
         tool=""
@@ -515,37 +538,6 @@ case "${mode}" in
         command_str=""
         command_str=$(printf '%s' "${json_input}" | jq -r '.tool_input.command // ""' 2>/dev/null) || true
 
-        # Inline AI context: detect the PID-scoped CCP_TASK_SUMMARY marker
-        # echoed by the main Claude session (injected via --append-system-prompt).
-        # The marker includes the ccp session PID (CCP_SESSION_PID), making it
-        # unique per session.  Source files on disk always contain the generic
-        # template string (without a real PID), so grep/cat of hook_runner.sh
-        # or bin/ccp can never produce a false match.
-        #
-        # Once captured, a marker file signals future invocations to skip scanning.
-        _inline_captured_file="${STATE_DIR:-/tmp}/inline_captured.${CCP_SESSION_PID:-$$}"
-        if [[ "${CCP_AI_CONTEXT_STRATEGY:-haiku}" == "inline" ]] && \
-           [[ -n "${CCP_CONTEXT_FILE:-}" ]] && \
-           [[ -n "${CCP_SESSION_PID:-}" ]] && \
-           [[ ! -f "${_inline_captured_file}" ]]; then
-            if [[ "${tool_response}" =~ CCP_TASK_SUMMARY_${CCP_SESSION_PID}:(.+) ]]; then
-                _inline_summary="${BASH_REMATCH[1]}"
-                _inline_summary=$(printf '%s' "${_inline_summary}" \
-                    | sed 's/^[[:space:]]*//' | sed 's/[[:space:]]*$//' \
-                    | sed "s/^['\"]//;s/['\"]$//")
-                if [[ -n "${_inline_summary}" ]]; then
-                    atomic_write "${CCP_CONTEXT_FILE}" "${_inline_summary}"
-                    touch "${_inline_captured_file}" 2>/dev/null || true
-                    _dbg_event "ai_summary" "summary=${_inline_summary}" "strategy=inline"
-                    _dbg "inline-summary: ${_inline_summary}"
-                fi
-            else
-                # Bash output didn't contain the inline marker — expected for most
-                # Bash calls, but useful for tracking whether a summary ever arrives.
-                _dbg_event "inline_marker_miss" "command=$(printf '%s' "${command_str}" | head -c 80)" "output_len=${#tool_response}"
-            fi
-        fi
-
         # Branch detection runs even without CCP_STATUS_FILE — skip status
         # detection only, not the entire handler.
         [[ -z "${CCP_STATUS_FILE:-}" && -z "${CCP_BRANCH_FILE:-}" ]] && exit 0
@@ -557,6 +549,18 @@ case "${mode}" in
             status="❌ Tests failed"
         elif [[ "${command_str}" =~ git[[:space:]]+commit && "${tool_response}" =~ ^\[ ]]; then
             status="💾 Committed"
+            # Capture commit subject line as task context — a human-written summary
+            if [[ -n "${CCP_CONTEXT_FILE:-}" ]]; then
+                _commit_subject=""
+                _commit_subject=$(printf '%s' "${tool_response}" \
+                    | head -1 \
+                    | sed 's/^\[[^]]*\][[:space:]]*//' \
+                    | head -c 80) || true
+                if [[ -n "${_commit_subject}" ]]; then
+                    atomic_write "${CCP_CONTEXT_FILE}" "${_commit_subject}"
+                    _dbg_event "context_set" "context=${_commit_subject}" "source=commit-message"
+                fi
+            fi
         elif [[ "${command_str}" =~ git[[:space:]]+push ]] && \
              [[ "${tool_response}" =~ (error:[[:space:]]+failed[[:space:]]+to[[:space:]]+push|!\ \[rejected\]|!\ \[remote\ rejected\]|ERROR:) ]]; then
             status="🐛 Push failed"
diff --git a/lib/session.sh b/lib/session.sh
index 0ab7776..afb2787 100644
--- a/lib/session.sh
+++ b/lib/session.sh
@@ -41,8 +41,8 @@ save_session() {
 
 # prune_dead_sessions: remove entries whose process is no longer running.
 # Rewrites SESSION_FILE in-place so all subsequent reads see only live sessions.
-# Also cleans up orphan state files (status, context, branch, monitor PID,
-# inline_captured) left behind by sessions killed without their EXIT trap.
+# Also cleans up orphan state files (status, context, branch, monitor PID)
+# left behind by sessions killed without their EXIT trap.
 prune_dead_sessions() {
     [[ ! -f "${SESSION_FILE}" ]] && return
 
@@ -63,7 +63,7 @@ prune_dead_sessions() {
     local f pid
     for f in "${STATE_DIR}"/status.*.txt "${STATE_DIR}"/context.*.txt \
              "${STATE_DIR}"/branch.*.txt "${STATE_DIR}"/agents.*.txt \
-             "${STATE_DIR}"/monitor.*.pid "${STATE_DIR}"/inline_captured.*; do
+             "${STATE_DIR}"/monitor.*.pid; do
         [[ -f "${f}" ]] || continue
         pid=$(basename "${f}" | grep -o '[0-9]\+') || continue
         if ! kill -0 "${pid}" 2>/dev/null; then
diff --git a/tests/test-suite.sh b/tests/test-suite.sh
index bb4108e..25638c0 100755
--- a/tests/test-suite.sh
+++ b/tests/test-suite.sh
@@ -483,6 +483,36 @@ result=$(printf '%s' '{"prompt":"Refactor the database layer"}' \
       bash "${LIB_DIR}/hook_runner.sh" user-prompt && cat "${TMP_CONTEXT}" 2>/dev/null || true)
 assert_contains "user-prompt: no trailing newline" "Refactor the database" "${result}"
 
+echo ""
+echo "user-prompt: context sanitization"
+
+TMP_STATUS="${STATE_DIR}/test-sanitize-status.txt"
+TMP_CONTEXT="${STATE_DIR}/test-sanitize-context.txt"
+
+# Pasted terminal content with venv prefix
+rm -f "${TMP_CONTEXT}" "${TMP_STATUS}"
+result=$(printf '%s' '{"prompt":"(venv) brianruggieri@Flexias-MacBook-Pro candidate-eval % fix the bug"}' \
+    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
+      bash "${LIB_DIR}/hook_runner.sh" user-prompt && cat "${TMP_CONTEXT}" 2>/dev/null || true)
+assert_not_contains "sanitize: venv prefix stripped" "(venv)" "${result}"
+assert_contains "sanitize: actual content preserved" "fix the bug" "${result}"
+
+# Pasted terminal with user@host prefix (no venv)
+rm -f "${TMP_CONTEXT}" "${TMP_STATUS}"
+result=$(printf '%s' '{"prompt":"user@hostname project % do the thing"}' \
+    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
+      bash "${LIB_DIR}/hook_runner.sh" user-prompt && cat "${TMP_CONTEXT}" 2>/dev/null || true)
+assert_not_contains "sanitize: user@host stripped" "user@hostname" "${result}"
+
+# Shell prompt character stripped
+rm -f "${TMP_CONTEXT}" "${TMP_STATUS}"
+result=$(printf '%s' '{"prompt":"$ npm test"}' \
+    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
+      bash "${LIB_DIR}/hook_runner.sh" user-prompt && cat "${TMP_CONTEXT}" 2>/dev/null || true)
+assert_equals "sanitize: $ prompt stripped" "npm test" "${result}"
+
+rm -f "${TMP_STATUS}" "${TMP_CONTEXT}"
+
 # stop handler clears status file
 printf '🧪 Testing' > "${TMP_STATUS}"
 echo '{}' \
@@ -491,6 +521,161 @@ echo '{}' \
 result=$(cat "${TMP_STATUS}" 2>/dev/null || true)
 assert_empty "stop empties status file"  "${result}"
 
+# ── Tests: stop-hook haiku summarization ──────────────────────────────────────
+
+echo ""
+echo "stop-hook haiku summarization"
+
+# Set up a mock claude binary that captures stdin and outputs a fixed summary
+MOCK_CLAUDE_DIR=$(mktemp -d)
+cat > "${MOCK_CLAUDE_DIR}/claude" <<'MOCK'
+#!/usr/bin/env bash
+# Mock claude binary — reads stdin, outputs a fixed summary
+cat > /dev/null  # consume stdin
+echo "Fixed Auth Login Flow"
+MOCK
+chmod +x "${MOCK_CLAUDE_DIR}/claude"
+
+# Each test uses a unique context file to avoid races with disowned background subprocesses
+HAIKU_CTX_1="${STATE_DIR}/haiku-ctx-1.txt"
+HAIKU_CTX_2="${STATE_DIR}/haiku-ctx-2.txt"
+HAIKU_CTX_3="${STATE_DIR}/haiku-ctx-3.txt"
+HAIKU_CTX_4="${STATE_DIR}/haiku-ctx-4.txt"
+HAIKU_CTX_5="${STATE_DIR}/haiku-ctx-5.txt"
+HAIKU_CTX_6="${STATE_DIR}/haiku-ctx-6.txt"
+
+# Test 1: stop hook with AI context enabled + last_assistant_message → haiku writes summary
+rm -f "${TMP_STATUS}" "${HAIKU_CTX_1}"
+printf '🧪 Testing' > "${TMP_STATUS}"
+printf '%s' '{"last_assistant_message":"I fixed the authentication flow by updating the login handler to properly validate tokens."}' \
+    | CLAUDECODE=1 \
+      CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${HAIKU_CTX_1}" \
+      CCP_ENABLE_AI_CONTEXT=true CCP_CLAUDE_BIN="${MOCK_CLAUDE_DIR}/claude" \
+      bash "${LIB_DIR}/hook_runner.sh" stop
+result=""
+for _ in {1..20}; do
+    [[ -f "${HAIKU_CTX_1}" ]] && result=$(cat "${HAIKU_CTX_1}" 2>/dev/null || true)
+    [[ -n "${result}" ]] && break
+    sleep 0.25
+done
+assert_equals "stop-haiku: writes summary to context file" "Fixed Auth Login Flow" "${result}"
+# Also verify status was still cleared (existing behavior preserved)
+status_result=$(cat "${TMP_STATUS}" 2>/dev/null || true)
+assert_empty "stop-haiku: status still cleared" "${status_result}"
+
+# Test 2: stop hook skips haiku when CCP_ENABLE_AI_CONTEXT is not true
+rm -f "${TMP_STATUS}" "${HAIKU_CTX_2}"
+printf '🧪 Testing' > "${TMP_STATUS}"
+printf '%s' '{"last_assistant_message":"I fixed the bug."}' \
+    | CLAUDECODE=1 \
+      CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${HAIKU_CTX_2}" \
+      CCP_ENABLE_AI_CONTEXT=false CCP_CLAUDE_BIN="${MOCK_CLAUDE_DIR}/claude" \
+      bash "${LIB_DIR}/hook_runner.sh" stop
+sleep 1
+result=$(cat "${HAIKU_CTX_2}" 2>/dev/null || true)
+assert_empty "stop-haiku: skips when AI context disabled" "${result}"
+
+# Test 3: stop hook skips haiku when last_assistant_message is empty
+rm -f "${TMP_STATUS}" "${HAIKU_CTX_3}"
+printf '🧪 Testing' > "${TMP_STATUS}"
+printf '%s' '{"last_assistant_message":""}' \
+    | CLAUDECODE=1 \
+      CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${HAIKU_CTX_3}" \
+      CCP_ENABLE_AI_CONTEXT=true CCP_CLAUDE_BIN="${MOCK_CLAUDE_DIR}/claude" \
+      bash "${LIB_DIR}/hook_runner.sh" stop
+sleep 1
+result=$(cat "${HAIKU_CTX_3}" 2>/dev/null || true)
+assert_empty "stop-haiku: skips when last_assistant_message is empty" "${result}"
+
+# Test 4: stop hook skips haiku when last_assistant_message field is missing from JSON
+rm -f "${TMP_STATUS}" "${HAIKU_CTX_4}"
+printf '🧪 Testing' > "${TMP_STATUS}"
+printf '%s' '{"some_other_field":"hello"}' \
+    | CLAUDECODE=1 \
+      CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${HAIKU_CTX_4}" \
+      CCP_ENABLE_AI_CONTEXT=true CCP_CLAUDE_BIN="${MOCK_CLAUDE_DIR}/claude" \
+      bash "${LIB_DIR}/hook_runner.sh" stop
+sleep 1
+result=$(cat "${HAIKU_CTX_4}" 2>/dev/null || true)
+assert_empty "stop-haiku: skips when last_assistant_message field missing" "${result}"
+
+# Test 5: messages >1000 chars are truncated to first 500 + "..." + last 500
+rm -f "${TMP_STATUS}" "${HAIKU_CTX_5}"
+printf '🧪 Testing' > "${TMP_STATUS}"
+# Build a mock claude that echoes back a snippet of what it received on stdin
+# so we can verify truncation happened
+MOCK_TRUNC_DIR=$(mktemp -d)
+cat > "${MOCK_TRUNC_DIR}/claude" <<'MOCK'
+#!/usr/bin/env bash
+# Mock claude binary — captures stdin content to a file for inspection
+input=$(cat)
+# Write the received input to a sidecar file for the test to inspect
+echo "${input}" > "${MOCK_TRUNC_DIR_SIDECAR}"
+echo "Truncation Test Summary"
+MOCK
+chmod +x "${MOCK_TRUNC_DIR}/claude"
+MOCK_SIDECAR="${MOCK_TRUNC_DIR}/captured_input.txt"
+
+# Build a 1200-char message: 600 A's + 600 B's
+long_msg=$(printf '%0600d' 0 | tr '0' 'A')$(printf '%0600d' 0 | tr '0' 'B')
+json_long=$(printf '{"last_assistant_message":"%s"}' "${long_msg}")
+printf '%s' "${json_long}" \
+    | CLAUDECODE=1 \
+      CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${HAIKU_CTX_5}" \
+      CCP_ENABLE_AI_CONTEXT=true CCP_CLAUDE_BIN="${MOCK_TRUNC_DIR}/claude" \
+      MOCK_TRUNC_DIR_SIDECAR="${MOCK_SIDECAR}" \
+      bash "${LIB_DIR}/hook_runner.sh" stop
+result=""
+for _ in {1..20}; do
+    [[ -f "${HAIKU_CTX_5}" ]] && result=$(cat "${HAIKU_CTX_5}" 2>/dev/null || true)
+    [[ -n "${result}" ]] && break
+    sleep 0.25
+done
+assert_equals "stop-haiku: truncated message produces summary" "Truncation Test Summary" "${result}"
+# Verify the captured input contains " ... " (truncation marker)
+if [[ -f "${MOCK_SIDECAR}" ]]; then
+    captured=$(cat "${MOCK_SIDECAR}" 2>/dev/null || true)
+    if [[ "${captured}" == *" ... "* ]]; then
+        pass "stop-haiku: truncated message contains ' ... ' marker"
+    else
+        fail "stop-haiku: truncated message contains ' ... ' marker" "captured input did not contain ' ... '"
+    fi
+else
+    fail "stop-haiku: truncated message contains ' ... ' marker" "sidecar file not found"
+fi
+rm -rf "${MOCK_TRUNC_DIR}"
+
+# Test 6: CLAUDECODE env var is unset in the haiku subprocess
+rm -f "${TMP_STATUS}" "${HAIKU_CTX_6}"
+printf '🧪 Testing' > "${TMP_STATUS}"
+MOCK_CLAUDECODE_DIR=$(mktemp -d)
+cat > "${MOCK_CLAUDECODE_DIR}/claude" <<'MOCK'
+#!/usr/bin/env bash
+cat > /dev/null
+if [[ -z "${CLAUDECODE:-}" ]]; then
+    echo "CLAUDECODE_ABSENT"
+else
+    echo "CLAUDECODE_PRESENT"
+fi
+MOCK
+chmod +x "${MOCK_CLAUDECODE_DIR}/claude"
+printf '%s' '{"last_assistant_message":"I fixed the bug by adding a null check."}' \
+    | CLAUDECODE=1 \
+      CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${HAIKU_CTX_6}" \
+      CCP_ENABLE_AI_CONTEXT=true CCP_CLAUDE_BIN="${MOCK_CLAUDECODE_DIR}/claude" \
+      bash "${LIB_DIR}/hook_runner.sh" stop
+result=""
+for _ in {1..20}; do
+    [[ -f "${HAIKU_CTX_6}" ]] && result=$(cat "${HAIKU_CTX_6}" 2>/dev/null || true)
+    [[ -n "${result}" ]] && break
+    sleep 0.25
+done
+assert_equals "stop-haiku: CLAUDECODE is unset in haiku subprocess" "CLAUDECODE_ABSENT" "${result}"
+rm -rf "${MOCK_CLAUDECODE_DIR}"
+
+rm -rf "${MOCK_CLAUDE_DIR}"
+rm -f "${TMP_STATUS}" "${TMP_CONTEXT}"
+
 # post-tool handler writes completion statuses from Bash output
 printf '🧪 Testing' > "${TMP_STATUS}"
 result=$(echo '{"tool_name":"Bash","tool_input":{"command":"npm test"},"tool_response":"3 tests passed"}' \
@@ -705,6 +890,24 @@ result=$(echo '{"tool_name":"Bash","tool_input":{"command":"git checkout main"},
 assert_equals "post-tool: branch detection fires without CCP_STATUS_FILE" "${expected_branch}" "${result}"
 rm -f "${TMP_BRANCH}"
 
+echo ""
+echo "post-tool: git commit message capture"
+
+TMP_STATUS="${STATE_DIR}/test-commit-ctx-status.txt"
+TMP_CONTEXT="${STATE_DIR}/test-commit-ctx-context.txt"
+
+# post-tool: git commit writes commit subject to context file
+rm -f "${TMP_STATUS}" "${TMP_CONTEXT}"
+result=$(echo '{"tool_name":"Bash","tool_input":{"command":"git commit -m \"Fix null check in parser\""},"tool_response":"[main abc1234] Fix null check in parser\n 1 file changed, 2 insertions(+)"}' \
+    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
+      bash "${LIB_DIR}/hook_runner.sh" post-tool && cat "${TMP_CONTEXT}" 2>/dev/null || true)
+assert_equals "commit: subject written to context file" "Fix null check in parser" "${result}"
+# Status should also be set
+status_result=$(cat "${TMP_STATUS}" 2>/dev/null || true)
+assert_equals "commit: status set to Committed" "💾 Committed" "${status_result}"
+
+rm -f "${TMP_STATUS}" "${TMP_CONTEXT}"
+
 # event handler: quiet (default) high-signal coverage
 result=$(echo '{"permission":"needed"}' \
     | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" CCP_STATUS_PROFILE=quiet \
@@ -985,161 +1188,6 @@ assert_equals "teardown preserves non-CCP hooks" "1" "${remaining}"
 rm -rf "${HOOKS_TMP_DIR}"
 rm -f "${TMP_STATUS}" "${TMP_CONTEXT}"
 
-# ── Tests: inline AI context (post-tool CCP_TASK_SUMMARY detection) ───────────
-
-echo ""
-echo "inline AI context (post-tool CCP_TASK_SUMMARY detection)"
-
-TMP_STATUS=$(mktemp)
-TMP_CONTEXT=$(mktemp)
-
-# Fixed PID used across all inline tests — mirrors the CCP_SESSION_PID export in bin/ccp
-TEST_PID=99999
-# Marker file created by hook_runner.sh after first inline capture
-INLINE_CAPTURED="${STATE_DIR}/inline_captured.${TEST_PID}"
-
-# post-tool: PID-scoped marker in echo output writes to context file
-rm -f "${TMP_CONTEXT}" "${TMP_STATUS}" "${INLINE_CAPTURED}"
-result=$(echo "{\"tool_name\":\"Bash\",\"tool_input\":{\"command\":\"echo CCP_TASK_SUMMARY_${TEST_PID}:Fix JWT Validation Bug\"},\"tool_response\":\"CCP_TASK_SUMMARY_${TEST_PID}:Fix JWT Validation Bug\"}" \
-    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
-      CCP_ENABLE_AI_CONTEXT=true CCP_AI_CONTEXT_STRATEGY=inline CCP_SESSION_PID="${TEST_PID}" \
-      bash "${LIB_DIR}/hook_runner.sh" post-tool && cat "${TMP_CONTEXT}" 2>/dev/null || true)
-assert_equals "inline: PID-scoped marker extracted" "Fix JWT Validation Bug" "${result}"
-
-# post-tool: marker with surrounding whitespace is trimmed
-rm -f "${TMP_CONTEXT}" "${TMP_STATUS}" "${INLINE_CAPTURED}"
-result=$(echo "{\"tool_name\":\"Bash\",\"tool_input\":{\"command\":\"echo CCP_TASK_SUMMARY_${TEST_PID}:  Refactor Auth Module  \"},\"tool_response\":\"CCP_TASK_SUMMARY_${TEST_PID}:  Refactor Auth Module  \"}" \
-    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
-      CCP_ENABLE_AI_CONTEXT=true CCP_AI_CONTEXT_STRATEGY=inline CCP_SESSION_PID="${TEST_PID}" \
-      bash "${LIB_DIR}/hook_runner.sh" post-tool && cat "${TMP_CONTEXT}" 2>/dev/null || true)
-assert_equals "inline: whitespace around summary trimmed" "Refactor Auth Module" "${result}"
-
-# post-tool: marker with quotes is cleaned
-rm -f "${TMP_CONTEXT}" "${TMP_STATUS}" "${INLINE_CAPTURED}"
-result=$(printf '{"tool_name":"Bash","tool_input":{"command":"echo CCP_TASK_SUMMARY_%s:Update Login UI"},"tool_response":"CCP_TASK_SUMMARY_%s:'\''Update Login UI'\''"}' \
-        "${TEST_PID}" "${TEST_PID}" \
-    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
-      CCP_ENABLE_AI_CONTEXT=true CCP_AI_CONTEXT_STRATEGY=inline CCP_SESSION_PID="${TEST_PID}" \
-      bash "${LIB_DIR}/hook_runner.sh" post-tool && cat "${TMP_CONTEXT}" 2>/dev/null || true)
-assert_equals "inline: quotes stripped from summary" "Update Login UI" "${result}"
-
-# post-tool: GENERIC (unscoped) marker does NOT match — this is the false-positive fix.
-# Source files on disk contain CCP_TASK_SUMMARY: without a PID; grep/cat of those
-# files must never overwrite a previously captured summary.
-rm -f "${TMP_CONTEXT}" "${TMP_STATUS}"
-printf 'previously captured summary' > "${TMP_CONTEXT}"
-result=$(echo '{"tool_name":"Bash","tool_input":{"command":"grep CCP_TASK_SUMMARY lib/hook_runner.sh"},"tool_response":"CCP_TASK_SUMMARY:Do Not Overwrite"}' \
-    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
-      CCP_ENABLE_AI_CONTEXT=true CCP_AI_CONTEXT_STRATEGY=inline CCP_SESSION_PID="${TEST_PID}" \
-      bash "${LIB_DIR}/hook_runner.sh" post-tool && cat "${TMP_CONTEXT}" 2>/dev/null || true)
-assert_equals "inline: unscoped marker ignored (false-positive guard)" "previously captured summary" "${result}"
-
-# post-tool: marker from a DIFFERENT PID does not match this session
-rm -f "${TMP_CONTEXT}" "${TMP_STATUS}"
-printf 'correct summary' > "${TMP_CONTEXT}"
-result=$(echo '{"tool_name":"Bash","tool_input":{"command":"echo CCP_TASK_SUMMARY_11111:Wrong Session"},"tool_response":"CCP_TASK_SUMMARY_11111:Wrong Session"}' \
-    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
-      CCP_ENABLE_AI_CONTEXT=true CCP_AI_CONTEXT_STRATEGY=inline CCP_SESSION_PID="${TEST_PID}" \
-      bash "${LIB_DIR}/hook_runner.sh" post-tool && cat "${TMP_CONTEXT}" 2>/dev/null || true)
-assert_equals "inline: wrong-PID marker ignored" "correct summary" "${result}"
-
-# post-tool: marker is IGNORED when strategy is not inline
-rm -f "${TMP_CONTEXT}" "${TMP_STATUS}"
-printf 'existing context' > "${TMP_CONTEXT}"
-result=$(echo "{\"tool_name\":\"Bash\",\"tool_input\":{\"command\":\"echo CCP_TASK_SUMMARY_${TEST_PID}:Do Not Overwrite\"},\"tool_response\":\"CCP_TASK_SUMMARY_${TEST_PID}:Do Not Overwrite\"}" \
-    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
-      CCP_ENABLE_AI_CONTEXT=true CCP_AI_CONTEXT_STRATEGY=haiku CCP_SESSION_PID="${TEST_PID}" \
-      bash "${LIB_DIR}/hook_runner.sh" post-tool && cat "${TMP_CONTEXT}" 2>/dev/null || true)
-assert_equals "inline: marker ignored when strategy is haiku" "existing context" "${result}"
-
-# post-tool: marker is IGNORED when AI context is disabled entirely.
-# bin/ccp never exports CCP_AI_CONTEXT_STRATEGY=inline when the feature is off,
-# so explicitly override to haiku here to simulate a clean non-inline environment.
-rm -f "${TMP_CONTEXT}" "${TMP_STATUS}"
-printf 'existing context' > "${TMP_CONTEXT}"
-result=$(echo "{\"tool_name\":\"Bash\",\"tool_input\":{\"command\":\"echo CCP_TASK_SUMMARY_${TEST_PID}:Do Not Overwrite\"},\"tool_response\":\"CCP_TASK_SUMMARY_${TEST_PID}:Do Not Overwrite\"}" \
-    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
-      CCP_AI_CONTEXT_STRATEGY=haiku CCP_SESSION_PID="${TEST_PID}" \
-      bash "${LIB_DIR}/hook_runner.sh" post-tool && cat "${TMP_CONTEXT}" 2>/dev/null || true)
-assert_equals "inline: marker ignored when AI context disabled" "existing context" "${result}"
-
-# post-tool: normal Bash output without marker does NOT touch context file
-rm -f "${TMP_CONTEXT}" "${TMP_STATUS}"
-printf 'existing context' > "${TMP_CONTEXT}"
-result=$(echo '{"tool_name":"Bash","tool_input":{"command":"ls -la"},"tool_response":"total 42\ndrwxr-xr-x 5 user staff 160 Jan  1 12:00 ."}' \
-    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
-      CCP_ENABLE_AI_CONTEXT=true CCP_AI_CONTEXT_STRATEGY=inline CCP_SESSION_PID="${TEST_PID}" \
-      bash "${LIB_DIR}/hook_runner.sh" post-tool && cat "${TMP_CONTEXT}" 2>/dev/null || true)
-assert_equals "inline: non-marker Bash output preserves context" "existing context" "${result}"
-
-# post-tool: PID-scoped summary + test pass — both context and status captured
-rm -f "${TMP_CONTEXT}" "${TMP_STATUS}" "${INLINE_CAPTURED}"
-result=$(echo "{\"tool_name\":\"Bash\",\"tool_input\":{\"command\":\"npm test\"},\"tool_response\":\"CCP_TASK_SUMMARY_${TEST_PID}:Fix Tests\n3 tests passed\"}" \
-    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
-      CCP_ENABLE_AI_CONTEXT=true CCP_AI_CONTEXT_STRATEGY=inline CCP_SESSION_PID="${TEST_PID}" \
-      bash "${LIB_DIR}/hook_runner.sh" post-tool && cat "${TMP_STATUS}" 2>/dev/null || true)
-assert_equals "inline: summary + test pass both captured (status)" "✅ Tests passed" "${result}"
-result=$(cat "${TMP_CONTEXT}" 2>/dev/null || true)
-assert_contains "inline: summary + test pass both captured (context)" "Fix Tests" "${result}"
-
-# user-prompt: inline strategy skips Haiku subprocess (just writes first-5-words)
-rm -f "${TMP_CONTEXT}" "${TMP_STATUS}"
-result=$(echo '{"prompt":"Fix the login bug in the auth module"}' \
-    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
-      CCP_ENABLE_AI_CONTEXT=true CCP_AI_CONTEXT_STRATEGY=inline \
-      bash "${LIB_DIR}/hook_runner.sh" user-prompt && cat "${TMP_CONTEXT}" 2>/dev/null || true)
-assert_contains "inline: user-prompt still writes first-5-words" "Fix the login bug" "${result}"
-
-# user-prompt: haiku strategy with AI context enabled (no claude binary in PATH)
-# just verify it writes first-5-words and doesn't fail (the background Haiku
-# subprocess will silently fail without claude binary — expected in tests)
-rm -f "${TMP_CONTEXT}" "${TMP_STATUS}"
-result=$(echo '{"prompt":"Refactor the database layer for performance"}' \
-    | CCP_STATUS_FILE="${TMP_STATUS}" CCP_CONTEXT_FILE="${TMP_CONTEXT}" \
-      CCP_ENABLE_AI_CONTEXT=true CCP_AI_CONTEXT_STRATEGY=haiku \
-      bash "${LIB_DIR}/hook_runner.sh" user-prompt && cat "${TMP_CONTEXT}" 2>/dev/null || true)
-assert_contains "haiku: user-prompt writes first-5-words" "Refactor the database layer" "${result}"
-
-rm -f "${TMP_STATUS}" "${TMP_CONTEXT}"
-
-# ── Tests: bin/ccp AI context strategy parsing ────────────────────────────────
-
-echo ""
-echo "bin/ccp --ai-context-strategy"
-
-BIN_CCP="${PROJECT_DIR}/bin/ccp"
-CLI_TMP_DIR=$(mktemp -d)
-CLI_STATE_DIR="${CLI_TMP_DIR}/state"
-mkdir -p "${CLI_STATE_DIR}"
-CLI_SESSION_FILE="${CLI_STATE_DIR}/sessions.json"
-echo '[]' > "${CLI_SESSION_FILE}"
-
-cli_exit=0
-CCP_CLAUDE_CMD=/usr/bin/true CCP_STATUS_PROFILE=quiet \
-STATE_DIR="${CLI_STATE_DIR}" SESSION_FILE="${CLI_SESSION_FILE}" \
-"${BIN_CCP}" --ai-context --ai-context-strategy inline --no-dynamic "strategy test" \
-    >/dev/null 2>/dev/null || cli_exit=$?
-assert_equals "bin/ccp: --ai-context-strategy inline succeeds" "0" "${cli_exit}"
-
-cli_exit=0
-CCP_CLAUDE_CMD=/usr/bin/true CCP_STATUS_PROFILE=quiet \
-STATE_DIR="${CLI_STATE_DIR}" SESSION_FILE="${CLI_SESSION_FILE}" \
-"${BIN_CCP}" --ai-context --ai-context-strategy haiku --no-dynamic "strategy test" \
-    >/dev/null 2>/dev/null || cli_exit=$?
-assert_equals "bin/ccp: --ai-context-strategy haiku succeeds" "0" "${cli_exit}"
-
-cli_exit=0
-cli_err="${CLI_TMP_DIR}/invalid-strategy.err"
-CCP_CLAUDE_CMD=/usr/bin/true CCP_STATUS_PROFILE=quiet \
-STATE_DIR="${CLI_STATE_DIR}" SESSION_FILE="${CLI_SESSION_FILE}" \
-"${BIN_CCP}" --ai-context --ai-context-strategy bogus --no-dynamic "strategy test" \
-    >/dev/null 2>"${cli_err}" || cli_exit=$?
-assert_equals "bin/ccp: invalid --ai-context-strategy exits 1" "1" "${cli_exit}"
-assert_contains "bin/ccp: invalid strategy emits clear error" \
-    "Invalid AI context strategy 'bogus'" "$(cat "${cli_err}" 2>/dev/null || true)"
-
-rm -rf "${CLI_TMP_DIR}"
-
 # ── Tests: bin/ccp status profile parsing ─────────────────────────────────────
 
 echo ""
@@ -1357,26 +1405,6 @@ mkdir -p "${CLI_STATE_DIR}"
 CLI_SESSION_FILE="${CLI_STATE_DIR}/sessions.json"
 echo '[]' > "${CLI_SESSION_FILE}"
 
-# --append-system-prompt should be filtered from Claude args display
-cli_out="${CLI_TMP_DIR}/startup-inline.out"
-CCP_CLAUDE_CMD=/usr/bin/true CCP_STATUS_PROFILE=quiet \
-STATE_DIR="${CLI_STATE_DIR}" SESSION_FILE="${CLI_SESSION_FILE}" \
-"${BIN_CCP}" --ai-context --ai-context-strategy inline --no-dynamic "messaging test" \
-    >"${cli_out}" 2>&1 || true
-assert_not_contains "bin/ccp: inline strategy hides --append-system-prompt from args display" \
-    "--append-system-prompt" "$(cat "${cli_out}" 2>/dev/null || true)"
-
-# Extra user-supplied args should still appear when --append-system-prompt is also injected
-cli_out2="${CLI_TMP_DIR}/startup-inline-extra.out"
-CCP_CLAUDE_CMD=/usr/bin/true CCP_STATUS_PROFILE=quiet \
-STATE_DIR="${CLI_STATE_DIR}" SESSION_FILE="${CLI_SESSION_FILE}" \
-"${BIN_CCP}" --ai-context --ai-context-strategy inline --no-dynamic "messaging test" -- --some-flag \
-    >"${cli_out2}" 2>&1 || true
-assert_contains "bin/ccp: extra args still shown when inline strategy active" \
-    "--some-flag" "$(cat "${cli_out2}" 2>/dev/null || true)"
-assert_not_contains "bin/ccp: --append-system-prompt still hidden when extra args present" \
-    "--append-system-prompt" "$(cat "${cli_out2}" 2>/dev/null || true)"
-
 # --goto: matching session emits "Resuming:" line
 CLI_GOTO_STATE_DIR="${CLI_TMP_DIR}/goto-state"
 mkdir -p "${CLI_GOTO_STATE_DIR}"