Run the board-write modes in CI

.github/workflows/ci.yml

@@ -46,6 +46,13 @@ jobs:
       # read-only guarantee would be asserted only by unit tests, never by a whole run.
       - run: npm run demo -- --agents
       - run: npm run demo -- --agents --provider anthropic
+      # The board agent as writer, replayed offline on both providers. Same reasoning as the two
+      # steps above, with more riding on it: this is the only mode in which a model reaches the
+      # tracker, so it is the only mode where a whole run demonstrates the governed write path
+      # refusing what the gates refuse. Off by default in normal use, so nothing else would exercise
+      # these recordings.
+      - run: npm run demo -- --agents --board-writes
+      - run: npm run demo -- --agents --board-writes --provider anthropic
       # Never --labels here: that path calls a judge model, and this job has no secrets.
       - run: npm run eval
 
@@ -57,13 +64,14 @@ jobs:
       # warning. A recording made against a prompt the code no longer sends is a reply that may not
       # be representative — the demo's entire claim is that it replays the real prompts.
       #
-      # If this fails: either re-record (`npm run record -- --all --agents --provider <p>`) or work
+      # If this fails: either re-record (`npm run record -- --all --agents [--board-writes] --provider <p>`) or work
       # out why the prompt moved. Do not silence it.
       - name: no cassette drift
         run: |
           set -uo pipefail
           drift=0
-          for mode in "" "--twice" "--provider anthropic" "--agents" "--agents --provider anthropic"; do
+          for mode in "" "--twice" "--provider anthropic" "--agents" "--agents --provider anthropic" \
+                      "--agents --board-writes" "--agents --board-writes --provider anthropic"; do
             # shellcheck disable=SC2086
             n=$(npm run demo --silent -- $mode 2>&1 | grep -c "recorded against a different prompt" || true)
             echo "demo ${mode:-(default)}: $n drifted cassette(s)"

AGENTS.md

@@ -136,6 +136,21 @@ happened to say on one day; if the feature needed a model to disagree in order t
 the demonstration would be the weather. But it does mean the honest claim is **"the path is proven by
 test, not by recording"** — and a reader who wants to see it fire should run the tests, not the demo.
 
+**The same is true of the board-write recordings, and for a reason worth recording.** Across all eight
+scenarios on both providers, `--board-writes` produces **zero refused writes**. Every write the board
+agent originates passes the gates.
+
+The first recorded attempt was not like that: it produced five refusals in a single scenario, all of
+them `unresolvable field(s): FINAL_DESC`. That was not the model failing the gate — `create_task`
+listed `description` as optional while the gate requires it, so the tool was lying about what a valid
+write looks like and the model believed it. Fixing the schema took the refusals to zero.
+
+Which leaves the same honest position as above: the governed write path is wired and exercised end to
+end, and **no shipped recording shows a gate refusing a write.** What proves it does is
+`governedTracker.test.ts` with scripted replies — an off-roster assignee, an unknown list key, a
+credential-touching title, a subtask whose parent history was never read, and an operation with no
+manifest form at all. Each is refused, and the inner adapter is asserted never to have seen it.
+
 ### This is what "authority to write" means
 
 The internal spec this repo was built from describes the Board agent as *"the orchestrator above the