diff --git a/docs/configuration/gga.md b/docs/configuration/gga.md
index 5f2c3940..72395135 100644
--- a/docs/configuration/gga.md
+++ b/docs/configuration/gga.md
@@ -15,6 +15,18 @@ Every gga review, in every repository, runs on the OpenAI Codex provider with
 | `~/.config/environment.d/50-gga-pin.conf` | Same provider and `PATH` for desktop-launched programs (IDE git hooks). |
 | Fish, Zsh, Bash startup | Export `GGA_PROVIDER=codex` and put the shim dir first on `PATH` when it exists. |
 
+## When Codex is unavailable
+
+The review is a safety net, not a wall. The block sets `STRICT_MODE="false"` in the global
+config, so a provider failure (for example `usage_limit_exceeded` when the Codex quota is used
+up) lets the commit through with gga's error on screen instead of blocking it, while a real
+`STATUS: FAILED` from the reviewer still blocks. gga never reports `PASSED` for a review that
+did not run. The quota is shared across models, so falling back to another model would not help.
+
+A repository whose own `.gga` sets `STRICT_MODE="true"` overrides the global value (project
+config is loaded after the global one) and keeps blocking during an outage: set it to `false`
+there, or commit with `git commit --no-verify` until the quota resets.
+
 ## Why a shim
 
 gga runs `codex exec "<prompt>"` without model options, so it would use the model
diff --git a/scripts/install-gga-pin.sh b/scripts/install-gga-pin.sh
index 468c42e8..42de6cc1 100755
--- a/scripts/install-gga-pin.sh
+++ b/scripts/install-gga-pin.sh
@@ -5,7 +5,7 @@
 # is kept outside it where possible and re-applied here idempotently:
 #   1. <gga>/bin/codex        shim that injects the model for gga's `codex exec`
 #   2. <gga>/pin.env          model and effort (created once, never overwritten)
-#   3. <gga>/config           marked block at the end: PROVIDER/GGA_PROVIDER=codex
+#   3. <gga>/config           marked block at the end: PROVIDER/GGA_PROVIDER=codex, STRICT_MODE=false
 #                             and the shim dir first on PATH
 #   4. environment.d/50-gga-pin.conf  the same for desktop-launched programs
 # One line is printed per changed file; nothing when everything is current.
@@ -96,6 +96,8 @@ ${BLOCK_BEGIN}
 # gentle-ai rewrites this file; re-run the installer or \`dreamcoder repair\` after it.
 PROVIDER="codex"
 GGA_PROVIDER="codex"
+# A real STATUS: FAILED still blocks; an unavailable provider (usage limit, network) does not.
+STRICT_MODE="false"
 case ":\${PATH}:" in *":${shim_dir}:"*) ;; *) export PATH="${shim_dir}:\${PATH}" ;; esac
 ${BLOCK_END}
 EOF
diff --git a/tests/shell/test_gga_pin.bats b/tests/shell/test_gga_pin.bats
index 0d173aec..041b1db3 100644
--- a/tests/shell/test_gga_pin.bats
+++ b/tests/shell/test_gga_pin.bats
@@ -88,6 +88,19 @@ shim() { PATH="${SHIM_DIR}:${REAL_BIN}:${PATH}" "$@"; }
     [ "$output" = "codex codex ${SHIM_DIR}" ]
 }
 
+# A review tool is a safety net, not a wall: a real `STATUS: FAILED` still blocks, but an
+# unavailable provider (usage limit, network) must not brick every commit. gentle-ai writes
+# STRICT_MODE="true" into the generated part of the file, so the block has to override it.
+@test "the block makes gga non-blocking when the provider is unavailable" {
+    gga_setup
+    gentle_ai_config
+    printf 'STRICT_MODE="true"\n' >>"${CONFIG}"
+    run installer
+    [ "$status" -eq 0 ]
+    run env -u GGA_PROVIDER bash -c 'source "$1"; printf "%s" "$STRICT_MODE"' _ "${CONFIG}"
+    [ "$output" = "false" ]
+}
+
 @test "sourcing the block twice does not duplicate the shim dir on PATH" {
     gga_setup
     installer >/dev/null