diff --git a/AGENTS.md b/AGENTS.md
index ea3a7af04c..e2a76755e2 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -11,7 +11,7 @@ agents work efficiently. Documentation is structured in three levels under
    This gives you setup, architecture, code map, conventions, workflows,
    interfaces, gotchas, and security.
 3. If a task needs more detail than L1 provides, follow links to L2 deep dives
-   in `docs/ai/L1/deep_dives/`. Load only the specific L2 file you need.
+   in `docs/ai/L1/L2/`. Load only the specific L2 file you need.
 
 ## Levels
 
@@ -51,9 +51,9 @@ agents work efficiently. Documentation is structured in three levels under
 | update docs   | code changed since last `last_reviewed` date  |
 | test docs     | verify docs give agents the right context     |
 
-For detailed procedures, read
-[progressive-disclosure-standard.md](docs/progressive-disclosure-standard.md)
-sections 6 (generate) and 7 (bootstrap).
+For detailed procedures, read `progressive-disclosure-standard.md`
+sections 6 (generate) and 7 (bootstrap) on the `docs/progressive-disclosure`
+branch (the standard is not checked into main).
 
 ## Working Areas
 
diff --git a/docs/ai/L1/01_setup.md b/docs/ai/L1/01_setup.md
index 5df1744df4..0160b7eb57 100644
--- a/docs/ai/L1/01_setup.md
+++ b/docs/ai/L1/01_setup.md
@@ -67,6 +67,12 @@ docker exec -d ten_agent_dev bash -c \
    task run > /tmp/task_run.log 2>&1"
 ```
 
+**Fresh worktree/checkout**: run `task install` FIRST (it includes the
+python-deps step). Running `install_python_deps.sh` alone on a fresh tree
+fails with `failed to build go app` because `ten_packages/system/` (including
+`ten_runtime_go`) is only populated by the install. Step 1 above is for
+refreshing deps after a container restart, when packages already exist.
+
 Typical choices:
 - `voice-assistant` for standard vendor iteration and demo graphs
 - `voice-assistant-advanced` for generated multi-graph setups
diff --git a/docs/ai/L1/05_workflows.md b/docs/ai/L1/05_workflows.md
index 7f992c4c94..3d40d5a8fa 100644
--- a/docs/ai/L1/05_workflows.md
+++ b/docs/ai/L1/05_workflows.md
@@ -164,6 +164,10 @@ catches a miss. For an extension change under
 # 1. Format + static checks (CI: task check, task lint)
 sudo docker exec ten_agent_dev bash -c \
   "cd /app && task format && task check && task lint"
+# task lint / task lint-extension in-container ALWAYS exits non-zero with
+# E0401 import-errors — the runtime imports only resolve in CI. Judge by
+# comparing against a known CI-green extension (e.g. xai_asr_python shows
+# the same E0401s); only warnings beyond that baseline are real findings.
 
 # 2. Standalone tests for the extension
 sudo docker exec ten_agent_dev bash -c \
diff --git a/docs/ai/L1/L2/extension_development.md b/docs/ai/L1/L2/extension_development.md
index 0ff01250a3..767e02040a 100644
--- a/docs/ai/L1/L2/extension_development.md
+++ b/docs/ai/L1/L2/extension_development.md
@@ -694,6 +694,8 @@ There are **10 tests** (1 excluded by the default test runner):
 | `test_dump`                 | Audio dump files created correctly                           |
 | `test_metrics`              | TTFW and TTLW metrics: positive, TTLW > TTFW                |
 | `test_audio_timestamp`      | start_ms and duration_ms accuracy                            |
+| `test_same_session_finalize_reconnect` | Audio sent after finalize still transcribes (fresh session if the vendor closes) |
+| `test_connection_status`    | connection_status_changed events — **allowlist-gated** (see below) |
 | `test_long_duration_stream` | **Excluded by default test runner** — 5+ min stream without timeout |
 
 ### Critical Pass Criteria
@@ -704,6 +706,29 @@ There are **10 tests** (1 excluded by the default test runner):
 - **Metrics**: TTFW > 0, TTLW > TTFW, both in milliseconds
 - **Audio format**: Accepts 16-bit PCM, 16kHz, mono, 320 bytes per frame
 
+### Guarder Contract Gotchas (learned the hard way)
+
+- **Timestamps** (`test_audio_timestamp`): every final needs `duration_ms > 0`,
+  and consecutive finals must not overlap (`prev.start_ms + prev.duration_ms
+  <= next.start_ms`). If the vendor returns no per-segment timing, synthesize
+  contiguous timestamps from the audio timeline: start = end of the previous
+  segment, end = total user audio sent when the final arrives.
+- **Invalid credentials must be NON-fatal** (`test_reconnection`): the test
+  feeds bad credentials and asserts every error in the window is non-fatal
+  (retry expected). Classify only HTTP 401/403 as immediately fatal; vendors
+  that reject bad keys another way (e.g. a websocket close code) must surface
+  a non-fatal error and let the reconnect manager escalate to FATAL at the
+  retry ceiling. Any fatal error inside the test window fails the test.
+- **connection_status tests are allowlist-gated**: the guarder only asserts
+  `connection_status_changed` sequences for extensions listed in
+  `_EXTENSIONS_WITH_CONNECTION_STATUS` (in
+  `integration_tests/asr_guarder/tests/test_connection_status.py`); other
+  extensions skip these two tests — expect `N passed, 2 skipped`.
+- **Capture the summary**: pipe guarder output to a log file
+  (`task asr-guarder-test ... > /tmp/guarder.log 2>&1; echo $?`) — piping to
+  `tail` reports the pipe's exit code, not pytest's, and the retry tracebacks
+  from `test_reconnection` push the summary line out of a short tail.
+
 ---
 
 ## AudioFrame Creation Pattern