The 'Record a workflow?' card appeared inconsistently because two intro prompt
lines still steered the agent the wrong way on an over-limit approve — 'review
... over-limit charges -> showPendingApprovals' and 'showApprovalFlow -> explain
how an over-limit charge gets cleared' — competing with the recall->offer block.
Scope both to browse/explain only and exclude them from the approve path, and
drop the agent temperature 0.3 -> 0 so it picks the same route every time.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Same lingering-button bug as the offer card: saveLearnedWorkflow and
awaitDashboardDemonstration only special-cased status 'inProgress', so after
clicking 'Save workflow' / 'I'm done' the card kept its button (confusing —
looked like nothing happened, though the save/respond went through). Add a
status 'complete' branch that collapses each to a static line, using the HITL
render's 'result' to show saved vs discarded / finished vs cancelled.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
On 'approve the over-limit charge' with no saved procedure, gpt-5.4-mini
sometimes called showApprovalFlow ('diagram of how to clear an over-limit
charge') or showPendingApprovals ('...including over-limit charges') instead of
recall_memory -> offerWorkflowRecording, so it never asked to record. Scope
those two tools to their real uses (explainer only when asked how it works;
queue only for reviewing pending) and explicitly exclude them from the
over-limit approve path. The approval-flow diagram stays as an informational
tool — just not a response to an approve request.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
After 'Start recording', offerWorkflowRecording resolved a bare 'started', so
gpt-5.4-mini often just SAID awaitDashboardDemonstration's 'go ahead and I'll
watch' line (from its description) instead of CALLING it — leaving the offer
card frozen with the Start-recording button and a confusing text reply, even
though recording (the vignette) had begun.
- Resolve a directive result telling the agent to immediately call
awaitDashboardDemonstration and not reply in prose (mirrors the working
awaitDashboardDemonstration -> saveLearnedWorkflow beat).
- Collapse the offer card to a static line once resolved (status 'complete')
so the button can't linger or be re-clicked; the live 'Recording your
workflow' card takes over.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Generalizes the row-action dropdown fix: every shadcn overlay (Select,
DropdownMenu, Popover, Tooltip) portals to <body> at z-50 and renders behind
the CopilotKit chat sidebar (z-1200) when opened in-chat — the policy-code
Select in the exception form hit this too. One globals.css rule lifts every
Radix popper wrapper to z-1300 (the file's existing 'above the chat panel'
value), fixing the whole class instead of per-component patches.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The 'More actions' menu portals to <body> at z-50, but the CopilotKit chat
sidebar is z-1200 — so the menu opened BEHIND the panel and was invisible
(clicking the three-dots appeared to do nothing). Lift this dropdown's content
to z-[1300] so it renders in front of the panel. Pairs with the earlier
pointer-events-auto fix; together the row's overflow menu works in-chat.
Verified: menu computed z-index is now 1300 vs panel 1200.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
showPendingApprovals is a display-only useComponent, which CopilotKit paints
pointer-events:none on the assistant message. The table is interactive
(Approve / Deny / File policy exception 'More actions' menu), so the inherited
none made every row action unclickable in the chat. Opt the card subtree back
into pointer events.
Verified live: the 'More actions' button's computed pointer-events flips
none -> auto with the fix.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Re-running the over-limit teach demo mutates the in-memory store (approve +
filed exception flip the charge overLimit→cleared). This endpoint re-seeds the
store in place so the $5,000 Google Ads / Marketing charge (t-1) returns to
pending/over-limit without a server restart. Gated off when NODE_ENV=production.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Drops the non-existent AIMock/MockServer fallback names (TS-flagged once aimock was
installed); LLMock/loadFixtureFile/validateFixtures are the real exports.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
AUTHORED + statically validated (playwright --list compiles spec+config, fixtures/
package JSON valid, launcher syntax OK). NOT yet green-verified — needs aimock
installed (pnpm i), the docker memory stack up, and the dev server in Intelligence
mode (multi-process; not runnable in the current sandbox). Each file carries a
'VERIFY ON FIRST GREEN RUN' checklist for the shakedown.
- e2e/memory-learning.spec.ts: seeds the procedure via REST (recall-half isolation
per the plan), drives a fresh thread, asserts recall->unlock with no recording
offer + the over-limit gate lifted. Save half stays HITL+LLM (drift smoke / manual).
- e2e/fixtures/memory-learning.fixtures.json: pins recall_memory -> openPolicyException
-> finalizePolicyException -> approveTransaction.
- e2e/aimock-server.mjs: aimock launcher (programmatic, CLI fallback documented).
- playwright.config.ts: webServer array (aimock + dev in Intelligence mode, OPENAI_BASE_URL->aimock).
- package.json: +@copilotkit/aimock devDep; test:self-learning -> the spec.
- remove scripts/self-learning-smoke.mjs (dead #192 distill path).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- README: replace dead sl-worker/knowledge/annotate Phase-C section with the
memory runbook (docker compose one-command stack, host-TEI override for Apple
Silicon, 715x ports, .env, cross-thread+cross-persona FOR-149 walkthrough,
aimock E2E + drift-smoke testing notes)
- scripts/memory-drift-smoke.mjs: non-gating real-LLM tripwire that the live model
still emits recall_memory on a fresh-thread over-limit request (save half is
HITL-gated; covered by the manual walkthrough + aimock E2E)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Save action now resolves a string the agent reads as status:saved + an explicit
save_memory(scope:project, kind:operational) instruction with the demonstrated
code, preserving the already-approved/don't-re-run guard. respond() takes a
string in this file, so the result is delivered as text the recall-first prompt
(Task 3) acts on.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Surfaced while verifying Task 0 against a live stack:
- minio-init: retry mc alias set until Docker DNS resolves (idempotent bucket create)
- embedder pluggable: MEMORY_EMBEDDINGS_URL overridable + bundled tei dependency required:false
(point at host/native TEI on RAM-constrained Apple Silicon where the emulated tei OOMs)
- deps default host ports remapped to 715x so a bare `docker compose up` coexists with a dev's Intelligence stack
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The chart components' descriptions steered the model to answer chart-shaped
questions BY rendering ("do NOT answer in plain text"), which it obeyed too
literally: "which policy is closest to its limit?" produced the right chart
and no answer. Every chart description now carries the same rule — the chart
replaces restating the raw numbers, not the answer itself; follow the render
with one or two grounded sentences. Pending-approvals card gets the matching
"point at what needs attention" line.
Verified live: the budget-usage pill now yields the chart plus "The Marketing
policy is closest to its limit…".
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Dismissal was persisted in sessionStorage, so after one engagement the
demo's opening beat — the copilot noticing breached charges before the
user types or clicks anything — never appeared again for the whole tab
session. Demo-wrong: every fresh load must open with it. Dismissal is now
per-page-load component state; within a load it still never re-nags (the
wrapper stays mounted across client-side navigation).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Click-around feedback fixes on the banking demo:
- In-chat pending approvals are now actually usable: the dashboard's 4-column
approval table is ~550px wide, so inside the chat card its Actions column
rendered past the card edge — visible but unclickable. New
PendingApprovalsChat stacks each charge as a chat-width card with labeled
Approve/Deny/File-exception actions, with exact behavioral parity
(over-limit gating, PolicyExceptionInline form, identical teach-mode
recordUserAction payloads and logStep narration).
- Suggestion bubbles never disappear: the full use-case catalog (12 pills —
the 4 self-learning arc beats plus charts, breakdowns, cash flow, approvals
explainer, report prep, PIN change, team invite) is registered with
available:"always", so the demo stays fully click-drivable after every
exchange. Panel widened 440→560px so pills flow two-per-row instead of
stacking.
- StatisticsChart (dashboard rail + spending-trend gen-UI) is a real chart
now: y-axis dollar gridlines, x-axis labels aligned to the plot area, and
pointer hover showing the exact month + amount with a guide line and
highlighted point. Still hand-rolled SVG, no charting dependency.
Validated with Playwright: approvals filed and approved entirely inside the
chat (Cleared state transition + queue drains), pills persist after
exchanges, tooltip renders on hover.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Make the banking demo pass the click-around test: every moment reachable by
clicking, zero typing, and the dead Analytics/Reports tabs replaced with real
views (FOR-190; claims FOR-178).
Additive modules under src/components/wow/:
- ProactiveNotice: on load, the copilot surfaces overnight policy breaches
unprompted and offers to walk through them (tap-to-accept -> existing
pending-approvals gen-UI). Session-scoped dismissal.
- ChartCard + AnalyticsView: the four brand charts (previously chat-only)
rendered as the dashboard's Analytics tab, each carrying contextual
conversation-starter pills grounded via the existing useAgentContext data.
- createReport tool + ReportsView: "prep the Q2 spend report" files a durable
artifact (summary + highlights + live charts) in the Reports tab through a
new /api/v1/reports REST route and store collection.
- useAskCopilot: shared open-panel + addMessage + runAgent helper (same path
a suggestion-pill click takes).
Correctness fixes found by a full Playwright click-through sweep:
- navigateToPageAndPerform: full-page reload tore down the chat panel mid-run
(conversation and in-flight operation lost); now client-side router.push.
Card operations also targeted /cards, which mirrors the dashboard and has no
card tools — they now land on / where the tools are registered.
- chat-inbox: CSS custom property cast so `next build` compiles again
(production build was failing on main, blocking deploy).
Validated end-to-end with Playwright against a live agent, including the full
teach -> demonstrate -> save -> recall arc in OSS mode.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Ports two fixes from the canonical oracle-cookbook demo into the oracle-agent-memory
showcase agent (server.py was byte-identical to the demo's pre-fix version):
1. Dangling tool-calls: booking conversationally calls the book_flight HITL tool,
which interrupts and emits an assistant tool_call awaiting the UI's Confirm/Cancel.
If the traveler sends another chat message instead, the unanswered tool_call made
the next turn 400 ("tool_call_ids did not have response messages"). _repair_dangling_tool_calls
synthesizes a "not completed" tool result for any dangling call before the history
reaches the graph. (Inverse of the duplicate-tool-block issue the history-replace
already handled — documented in docs/known-issues/agentspec-multiturn-toolcall-correlation.md.)
2. HTML-escaped persistence: the agentspec exporter HTML-escapes streamed deltas
(& < > -> & < >); the SSE generator persisted them raw, so assistant
replies were stored in Oracle Agent Memory as e.g. "fares < $700".
_clean_assistant_text html.unescapes the assembled text before persisting.
Both are reproduced + verified in oracle-cookbook (unit tests + in-browser); the ported
functions are byte-identical to the verified demo code.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Flag-gated (LANGGRAPH_CHECKPOINTER=oracle, default memory) AsyncOracleSaver from
langgraph-oracledb for durable per-thread LangGraph graph state in Oracle,
complementing oracleagentmemory. Default-safe (in-memory fallback). Mirrors
jerelvelarde/oracle-cookbook#4. Draft, stacked on #5563.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Sync the three travel-concierge UX fixes from the source repo
(jerelvelarde/oracle-cookbook#3) into the showcase:
- agent/concierge/server.py: persist memory in a serialized background task so
the SSE stream closes at RUN_FINISHED instead of blocking ~2-13s on memory
extraction + reconciliation (the post-generation loading lag). Adds a FastAPI
lifespan drain + background-task failure surfacing.
- frontend/.../FlightOptions.tsx: "Select this flight" now drives confirm -> book
entirely client-side, so the booking confirm card renders inline in view
instead of off-screen via a fresh agent turn. The conversational book_flight
HITL path is unchanged.
- frontend/e2e: add a covering test for the inline booking flow; rename
sendAndPersist -> sendAndAwaitRun (stream-close is no longer a "memory written"
signal now that persistence is backgrounded).
oxfmt --check + tsc --noEmit clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Adds a durable persistence layer for @copilotkit/bot, replacing the
in-memory-only ActionStore with a pluggable StateStore.
- StateStore interface (kv/list/lock/dedup/queue) with a shared
conformance suite; MemoryStore default plus @copilotkit/bot-store-redis
and @copilotkit/bot-store-postgres backends.
- createBot({ store }): typed per-thread state via Standard Schema,
action snapshots persisted through the store, per-conversation turn
lock (onLockConflict drop|force), and inbound-event dedup keyed on a
stable eventId. ActionStore is kept as a deprecated alias.
- Cross-platform transcripts (bot.transcripts + identity resolver) with
age-bounded retention (prune on append + filter on read), and
runAgent({ transcript: true }) to auto-inject history and capture the
reply.
- createBot({ components }) re-registers components so durable actions
re-fire after a restart; restart-durability demo in examples/slack.
- Dedup is marked seen only after the turn lock is acquired, so a turn
dropped on lock-conflict does not burn its eventId (no lost retries).
- Release lockstep: bot-store-redis/postgres version with bot + bot-ui.
The standalone showcase installs with `npm install` (Nixpacks does this
automatically when there's no lockfile; the README already uses npm install).
Removing the 16.6k-line lockfile also clears the fork-pr-monitor security
alert, which flagged it as a "large added file." Verified: a fresh
`npm install` + `next build` succeeds with no committed lockfile.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
build-and-push.sh defaulted IMAGE to a personal GHCR namespace; use an
OWNER placeholder (still IMAGE-overridable) so the showcase ships no
personal/private registry path. Resolves the pre-merge follow-up.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Synced from the cookbook dev repo (verified live on Railway + Playwright):
- db/init/01-create-user.sql: create the `cookbook` user on a dedicated
ASSM tablespace (cookbook_ts) as its DEFAULT, instead of inheriting
SYSTEM/MSSM (FREEPDB1's default on the Free :latest-lite image). JSON and
VECTOR columns are SecureFile-backed and cannot live in an MSSM
tablespace, so the old user failed schema creation with ORA-43853 and
memory never persisted. Self-heals a user stranded on SYSTEM
(drop+recreate); leaves a working ASSM user (local :latest USERS) alone.
- frontend: a new thread is named from its OWN first message, not the
previous thread's. ThreadTitler now titles off the agent's message/run
events gated on agent.threadId === activeThreadId, replacing a useEffect
keyed on activeThreadId that read the shared, agentId-scoped transcript
while it still held the prior thread's messages during a switch. Adds a
Playwright regression test (e2e/thread-title.spec.ts).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The opengenui route used the nonexistent model id "openai/gpt-5.2", causing
every chat turn to error. Switch to "openai/gpt-4o", the verified-real id used
by the sibling showcase/shell copilotkit route.
The example's API routes import @copilotkit/runtime/v2 (a server-only package),
but next.config.ts lacked serverExternalPackages, so Next.js attempted to bundle
it. Add serverExternalPackages: ["@copilotkit/runtime"] (base package name covers
the /v2 subpath), mirroring showcase/shell/next.config.ts.
Runnable companion to the "Build an Agentic Travel App with Oracle Agent
Memory, Agent Spec, and CopilotKit" cookbook recipe: a portable Oracle Agent
Spec agent on LangGraph over AG-UI, long-term memory on Oracle AI Database,
and a CopilotKit V2 frontend with generative UI + human-in-the-loop.
- agent/ Python (uv) Agent Spec agent + FastAPI AG-UI server
- frontend/ Next.js CopilotKit V2 chat (flight cards, recall chip, HITL booking)
- db/ Oracle AI Database (Free image) + cookbook-user init
- docker-compose.yml for local Oracle; per-service railway.json for deploy
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add display-only generative-UI the copilot renders on request:
- showPendingApprovals — the dashboard's interactive approval table, in chat
- showSpendingTrend, showBudgetUsage, showSpendBreakdown, showIncomeVsExpenses —
brand-styled charts (dependency-free, hand-rolled SVG/CSS)
- showApprovalFlow — a diagram of how an over-limit charge gets cleared
Registered globally in copilot-context (so they work on any route) and routed
via the agent prompt.
- Cards page: the credit-card grid compressed each card (wrapping the holder /
valid-thru text) when the chat panel narrowed the content area. It's now a
horizontal-scroll row — each card holds a 300px minimum and grows to fill on
wide screens, scrolling instead of squishing.
- overflow-x:auto forces overflow-y:auto, which clipped the cards' soft drop
shadows; matching negative-margin/padding pairs give the shadows room inside
the scroll viewport without moving the cards.
- Chat: space stacked HITL/action cards so consecutive cards no longer share an
edge (CopilotKit stacks them with no row-gap).
Adds a cookbook recipe + runnable showcase that gives the Built-in Agent
OAuth-backed Arcade tools (Gmail, Google News) and renders Arcade's one-time
authorization step as a generative-UI "Connect" card in the chat.
- docs: showcase/shell-docs/src/content/docs/cookbook/arcade.mdx (+ meta.json, index card)
- app: examples/showcases/arcade-tools (Next.js App Router, single-route runtime)
Co-authored-by: Cursor <cursoragent@cursor.com>
useCreditCards() kept per-instance React state and was called independently by
the dashboard page and by copilot-context (where the chat's approve / finalize /
open-exception tools live). A mutation made through one instance refetched only
itself — so when the agent approved an over-limit charge in chat, the dashboard
pending table kept showing it as pending until a manual reload.
Add a module-level revalidation bus: each useCreditCards() instance registers a
refetch callback, and every mutation calls notifyDataChanged() to fan a refetch
out to all live instances. The dashboard now reflects agent-driven approvals
immediately (verified: recall-approve in chat drops the charge from the pending
table with no reload).
Adds a "What each mode actually recalls" subsection to the demo README: in OSS
mode the taught workflow is recalled only within the same conversation (the
saved procedure is echoed back into that thread), so a brand-new chat won't know
it — that's expected. Cross-conversation persistence is what the external
Intelligence backend provides. Names the symptom explicitly so reviewers aren't
surprised when a new conversation "doesn't know" the workflow in OSS mode.
Rework the dashboard's Pending approval view and fix two approval-gate UX bugs
on the banking demo (PR #5266):
- Table layout: replace the center-stacked per-row card with a scannable table
(Merchant / Amount / Policy / Actions). Actions are check / x icon buttons plus
a "more actions" overflow menu holding File policy exception. Status is its own
column (Over limit / Cleared / Within limit), and Approve is gated until the
charge is actually clearable.
- Fix the table shrinking when the more-actions menu opens: the Radix menu is
modal by default and engaged react-remove-scroll, whose scrollbar compensation
reflowed the table. Set modal={false} (a row menu needn't be modal) and add
whitespace-nowrap to the status badges.
- Fix "cannot approve after filing an exception": the inline card offered all
codes, including non-justifying ones that set activeExceptionId (flipping the
row to Cleared) but never lift the server gate, so the approve 422'd silently.
The card exists to clear an over-limit charge, so it now offers only justifying
codes. The gate itself is unchanged.
Verified live in OSS dev: file (justifying) -> Cleared -> approve succeeds; the
menu opens without reflow; lint + build green.
The recorder was a floating HUD overlay, which read as a different UI element
from the in-chat cards (e.g. "Save this workflow?"). Fold the live step feed
into the demonstration card itself: the `awaitDashboardDemonstration` card is
now titled "Recording your workflow" (with a REC pulse), keeps the
non-directional copy, and embeds a reactive `<RecordingSteps/>` child that
narrates each captured action inside the card. Removes the floating panel and
its CSS. Same card chrome as the other cards, so it reads consistently in the
conversation.
`RecordingSteps` subscribes to the recording context itself (not via the host
card's render closure), so it updates live as each step is logged without a
stale-closure dep.
Address review feedback on the banking self-learning "teach a workflow" loop
(PR #5266):
- Recorder HUD: a floating "Recording your workflow" panel narrates each officer
action live (Opened Dashboard -> Opened Transactions -> Opened Pending approval
-> Opened the exception form -> Filed the policy exception -> Approved the
charge), driven by logStep on the nav / tab / file-exception / approve call
sites. New recording-feed.tsx; steps + logStep added to recording-context.
- Non-directional demonstration: the await card is retitled "Show me how" and no
longer lists the steps ("go ahead and do it yourself now and I'll watch and
learn"); the agent's spoken handoff is likewise generic.
- Fix the "nothing happens after I'm done" stall: the model sometimes asked
"should I save this?" in prose instead of calling saveLearnedWorkflow, leaving
no Save card to click. The await tool-result is now directive (call
saveLearnedWorkflow; the card is how you ask), reinforced in the prompt.
- Harden the ending: after saving, the agent treats the demonstrated charge as
already cleared and waits, instead of re-running the freshly-saved procedure
on it.
Verified end-to-end in OSS dev (taught Google Ads, recalled AWS); lint + build
green.
Rework the banking demo's teach-a-workflow loop so the officer demonstrates the
over-limit unlock on the actual dashboard instead of an inline chat card, and so
the first over-limit approve request no longer shows an approval card that fails.
When asked to approve an over-limit charge it has no saved procedure for, the
agent now declines and offers to record (no approval card). The officer opens the
new Dashboard -> Transactions -> Pending approval view, files a policy exception
and approves the charge there; a waiting card holds the chat until they click
"I'm done". The agent then saves the procedure and applies it itself to a
different over-limit charge.
Move the teach/recall HITL tools (offerWorkflowRecording,
awaitDashboardDemonstration, saveLearnedWorkflow, openPolicyException,
finalizePolicyException, approveTransaction) and the agent data/permission
readables into the global CopilotContext. A route-scoped registration unmounts
when the officer navigates to the dashboard, which drops the in-progress card and
prevents the followUp from firing; global registration survives navigation and
renders on every route.
The demonstrated exception code is captured via the recording context and handed
to the Save step, so the saved procedure names the exact code used. The agent
prompt is updated for the decline+offer and dashboard handoff, and still never
spells out the unlock or names a justifying code.
Verified end-to-end in OSS dev (lint + build green, gate smoke 3/3): Beat 1 shows
no card, the dashboard demonstration clears the Google Ads charge, and recall
clears the AWS charge via the learned procedure.
The self-learning recorder POSTs to the annotate endpoint, which only exists with an
Intelligence backend; in OSS mode it returns 422. Call sites logged that rejection with
console.error, which Next.js 16 surfaces as a full-screen dev error overlay mid-demo even
though the failure is expected and harmless.
Swallow the failure in the recorder seam (useRecordUserActionInCurrentThread): catch it and
log quietly via console.debug instead of letting it reject. Recording stays best-effort — a
no-op without an Intelligence backend, and unchanged (records normally) with one.
Drive the FOR-137 self-learning story as an agent-orchestrated, narrated loop. When an
over-limit approval is rejected, the agent offers to record a workflow; the officer
demonstrates by filing a policy exception; the agent summarizes and saves the procedure;
then it applies that procedure itself to a different over-limit charge. Same-session recall
works by echoing the learned procedure back into the thread.
page.tsx: three new human-in-the-loop tools (offerWorkflowRecording,
recordExceptionDemonstration, saveLearnedWorkflow) plus a transactions agent-readable so the
agent resolves a charge the user names to the right id instead of guessing.
openPolicyException now returns the new exception id, and the agent-driven exception tools
are followUp:true so the recall chain (open then finalize then approve) does not stall.
route.ts: TEACH & RECALL prompt rules and an ACTION DISCIPLINE carve-out. The prompt does
not restate the unlock procedure, preserving the learning invariant.
policy-exception-inline.tsx: surface the demonstrated exception code via onFiled(code).
scripts/over-limit-gate-smoke.mjs: regression guard proving only a finalized
justifying-code exception lifts the policy-limit gate.
Verified end-to-end in OSS dev mode (lint and build green): the demonstration clears the
Google Ads charge and recall clears the AWS charge with a single successful approve.
Rewire the before-first-message suggestion pills to drive the FOR-137
self-learning story: (1) the teachable over-limit ask, (2) surface the pending
charges so the officer can demonstrate the unlock, (3) recall on a different
over-limit charge on a fresh thread. Titles stay symptom-only so they do not
hint at the exception path the agent is meant to learn on its own.
Maxim