The "Sort out my offsite expenses" beat had three problems on stage: the
harness console was a black slab in a light-mode transcript, the tool
activity grew a stack of finished steps that pushed the report card off
the screen, and the run took a full minute.
Console (banking skin):
- Every colour is now a semantic token, so the pane follows the app into
dark or light instead of being hardcoded dark. It was `bg-ink` with
`text-white/45`-style overlays, which only ever looked right in one mode.
- Collapsed by default. The status strip still carries the live state, so
the run reads as alive while closed.
- Still the FULL log when open. It is the detail view, and windowing it as
well left two lines and nowhere to read the rest.
Tool activity (shell, all skins):
- Rolls to the last two lines; older ones are REMOVED, not collapsed.
- Registration is a layout effect, and that is load-bearing. A new line
renders before it is registered, and registering is what evicts the
oldest, so with a passive effect the browser painted the in-between
state: three rows for one frame on every tool call, and again when the
end-of-run MESSAGES_SNAPSHOT remounts every line at once. Measured per
animation frame over a full run: 12,734 frames, never more than two.
- A shared registry rather than something simpler because CopilotKit
renders one component per tool call and owns the container. CSS
`:nth-last-child` needs siblings and the lines had one parent each;
mount-order counters drift across the snapshot remount.
Agent (run time 1m 0s -> ~45s):
- Research is gated on the offsite dates. A charge dated outside the
window is settled by its date whatever the merchant turns out to be, so
half the researcher dispatches were buying nothing. Travel on the
adjacent days is still kept in scope.
- Filings go out in one command instead of one curl per row. The
researchers already ran concurrently, so the serial per-row round-trips
through the model were most of the wall clock, not the research.
- Fetch and verify are one command; there is nothing to decide between
the halves.
- Analyst reasoning effort defaults to medium, overridable with
BANKING_EXPENSE_EFFORT.
Filing is idempotent, which the batching made necessary: the script got
run twice and every charge was filed twice, so the report card claimed six
filings while the ledger held twelve. The script now writes `filed.json`
and exits early if it exists. That marker is cleared once per run, because
the workspace is a fixed directory shared by every run and a stale marker
would convince the next demo it had already filed and post nothing at all
— the same bug wearing the opposite mask, and a quieter one, since a run
that files nothing still writes a confident report.
Deduping server-side on merchant+amount would have been wrong: Hotel
Verrano legitimately appears twice at the identical 318.55 for the two
nights of the offsite.
Does this change make anything in .claude/skills/reskin/ wrong,
incomplete, or misleading? No. Nothing there documents the harness
console, the tool-activity renderer, or the analyst prompt; its only
matches for "console" are `console.error` in unrelated template code.
Verified on the real path against a live Intelligence stack: six rows
filed, all status=pending, stable across 60s of polling, and present in
the Pending Approval queue with their notes and approve/decline actions.
## Summary
- add an independently runnable Claude Managed Agents finance-assistant
example
- add a cookbook recipe that explains the CopilotKit runtime,
managed-session mapping, and tool rendering flow
- add the recipe to cookbook navigation, the overview grid, sidebar icon
mapping, and render coverage
- use the real Claude vector mark for the cookbook instead of the
text-placeholder SDK asset
- register the example's Vite configuration and managed-agent model with
the repository CI allowlists
- include a compact architecture diagram and links to the relevant
rendering and CopilotKit Intelligence documentation
- disable Claude's complete built-in toolset and expose only the scoped
`show_growth_projection` runtime tool
- make the provisioning model configurable through `ANTHROPIC_MODEL`,
defaulting to `claude-fable-5`
- bound CopilotKit request bodies to 256 KB and managed-agent turns to
90 seconds, while relying on the adapter's per-thread serialization
- cap public run traffic at 20 provider-like attempts per client IP per
minute and 2,000 successful starts per process per 24-hour window
- restrict browser runtime requests with an exact Origin allowlist that
supports same-origin or separately hosted frontends, and restrict iframe
parents with CSP `frame-ancestors`
- validate persisted managed-agent IDs at startup so malformed local
configuration fails immediately
- publish the interactive example on Railway and embed the live
deployment in the cookbook
- align the demo with the existing cookbook chat styling and show the
`Project monthly investing` starter on first load
## Demo
## Why
This gives developers a focused example of connecting CopilotKit to
Anthropic Claude Managed Agents without the extra surface area of a
larger analyst application. The recipe follows the existing cookbook
structure and keeps the live demo compact enough for the standard
cookbook pane. Its managed environment has no outbound network or
package-manager access, and its agent cannot use bash, filesystem,
search, or fetch tools.
The request, turn, per-IP, and process-wide limits bound public demo
traffic without adding authentication or user friction. The traffic
counters are intentionally in memory, reset on process restart, and are
not shared across replicas, so the dedicated Anthropic workspace spend
limit remains the durable cost backstop. The exact-Origin browser check
reduces drive-by use but is explicitly documented as a control rather
than authentication. The model override allows operators to select a
lower-cost supported model during provisioning without editing source
code.
## Validation
- scoped formatting: passed
- scoped lint: 0 warnings, 0 errors
- shell-docs typecheck: passed
- standalone example typecheck: passed
- docs render tests: 26/26 passed
- standalone example tests: 23/23 passed
- shell-docs tests: 375/375 passed
- shell-docs production build: passed (222/222 pages)
- standalone example production build: passed
- standalone npm lockfile validation: passed
- build-config allowlist validator: passed
- docs model-name validator: passed
- exact-Origin regression coverage for run requests plus headerless
same-origin runtime discovery: passed
- malformed persisted agent-ID regression coverage: passed
- live Railway root and iframe CSP: passed
- live Railway runtime discovery, exact welcome copy, and first-load
starter pill: passed
- live three-turn AG-UI managed-agent run with `show_growth_projection`:
passed
- cookbook verified in the browser at desktop and narrow widths with no
console errors or horizontal overflow
## What does this PR do?
Closes the naming and documentation half of
[OSS-881](https://linear.app/copilotkit/issue/OSS-881). Paired with
CopilotKit/Intelligence#890, which adds `copilotkit verify` and tightens
the evaluation rubric.
### 1. One name for the Intelligence key
**Three** names for one value were live in CopilotKit's own
documentation, and following the wrong one with a CLI-provisioned
project yields an undefined key:
| Name | Where | Code readers |
| --- | --- | --- |
| `INTELLIGENCE_API_KEY` | what `copilotkit project select` writes; all
34 integration examples; the docs site | 34 |
| `COPILOTKIT_INTELLIGENCE_API_KEY` | 7 Channels package READMEs +
packaged skills | **0** |
| `COPILOTKIT_API_KEY` | `examples/slack`, `examples/teams`, and the
TSDoc on `CopilotKitIntelligence` itself | 2 |
`INTELLIGENCE_API_KEY` wins — it is the name the CLI provisions, and
changing it would break every scaffolded project in the wild.
- `COPILOTKIT_INTELLIGENCE_API_KEY` is **retired outright**. Nothing
ever read it, so there is nothing to keep compatible.
- `COPILOTKIT_API_KEY` stays **readable as a deprecated alias** in the
two examples that consume it, so an existing `.env` keeps working, and
is documented as deprecated everywhere it appears.
The third name was the worst placed: it was in the TSDoc on
`CopilotKitIntelligence`, which is what an IDE shows on hover.
This was not only untidy. The CLI's own `channels-preflight` accepts
`INTELLIGENCE_API_KEY` or `COPILOTKIT_API_KEY` — **not**
`COPILOTKIT_INTELLIGENCE_API_KEY`, the name the Channels READMEs told
people to set. So following a Channels README verbatim made `copilotkit
channels` warn that no runtime API key was present while the key sat
visibly in `.env`. After this PR the documented name is one preflight
accepts.
> [!NOTE]
> `NEXT_PUBLIC_COPILOTKIT_API_KEY` is a **different value** — the legacy
Copilot Cloud public key — and is deliberately left alone.
### 2. A real defect, not just naming skew
`skills/runtime/references/intelligence-mode.md` documented
`organizationId` as a `CopilotKitIntelligence` option, sourced from two
further env names (`COPILOTKIT_INTELLIGENCE_ORG_ID`,
`COPILOTKIT_ORG_ID`).
`CopilotKitIntelligenceConfig` has no such field — the copy-pasteable
sample it appeared in **would not compile**. Removed from the samples,
and the prose telling readers to fetch a value for it corrected. That
file is the only place those two names ever existed, which is very
likely why the failing validation run reported that "the runtime reads
`COPILOTKIT_INTELLIGENCE_API_KEY` and `COPILOTKIT_INTELLIGENCE_ORG_ID`".
### 3. Publish the Intelligence wiring
The wiring instructions existed only inside
`node_modules/@copilotkit/runtime/skills/`, and the only docs pages
mentioning `CopilotKitIntelligence` at all were the two Channels
frontends — so a developer on the plain web path had no page to reach it
from.
Adds **`/premium/connect-your-runtime`**: the wiring itself, how to
confirm the credential is actually consumed, the self-hosted
both-URLs-or-neither rule, and a troubleshooting table. Linked into both
navs, and the skills reference now points at the published page.
### 4. A guard so it cannot drift back
`scripts/validate-intelligence-env-names.ts` (`pnpm
check:intelligence-env-names`), wired to lefthook and a new workflow.
The workflow is **intentionally unfiltered**. The two workflows that
would otherwise cover this both filter: `plugin-skills-check` by
`paths:`, and `static/quality` by `paths-ignore: examples/**` — which is
exactly where the deprecated alias lives. Scoping the job would re-open
the hole it exists to close. Legitimate alias sites live in
`ALIAS_ALLOWLIST`.
## Related PRs and Issues
- [OSS-881](https://linear.app/copilotkit/issue/OSS-881) — needs
**both** PRs; neither closes it alone
- CopilotKit/Intelligence#890 — items 1 and 4 (`copilotkit verify` +
rubric contract 1.3.0)
## Verification
- Full lefthook pre-commit ran green: `check-plugin-skills`, `lint-fix`,
the new `check-intelligence-env-names`, and `test`/`publint`/`attw`
across **25 projects**.
- `examples/slack` `managed.test.ts` extended to cover **both** the
canonical name and the alias fallback, and proven non-vacuous — removing
the fallback turns the new test red.
- The drift guard proven non-vacuous the same way: reintroducing a
retired name fails it, exit 1.
- `oxfmt` and `oxlint` clean on every file touched (0 errors).
## Checklist
- [x] I have read the [Contribution
Guide](https://github.com/copilotkit/copilotkit/blob/master/CONTRIBUTING.md)
- [x] If the PR changes or adds functionality, I have updated the
relevant documentation
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Replaces the Codex CLI harness with **LangChain deep agents (Python)**:
sandboxed shell execution, parallel research subagents, no external
harness binary, our own API key.
All ten beats have been walked in a browser, and the demo now has its
own CI job. **Ready for review.**
## What banking's agent is now
```
banking gpt-5.4, temp 0
│ banking's own prompt · the browser's frontend tools · Intelligence
│ memory tools · render_report for the canvas
└─ expense-analyst gpt-5.6-sol, reasoning_effort=high
│ sandboxed shell (LocalShellBackend) · submit_expense_report
└─ merchant-researcher gpt-5.4, one per merchant
search_merchant (Tavily)
```
Six skins still run in-process as `BuiltInAgent`s. `banking` is an
`HttpAgent` pointed at `agent/`.
Each level exists because something about it must differ from its parent
— prompt, model, reasoning effort, tool set. A single agent had nowhere
to put any of that.
### Why the whole agent moved, not just the expense beat
The obvious design was a second agent id for the long-running beat,
leaving banking's `BuiltInAgent` alone. That was built first, and it
does not survive the requirement *"start the analysis, switch threads,
run other pills, come back to it."*
**Threads are scoped per agent.** `listThreads` takes `agentId` as a
required parameter, and measured against the running app the two lists
were disjoint — `banking` 46 threads, `banking-expenses` 10. The v2
runtime has no handoff mechanism, and `defineTool`'s `execute` takes
`(args)` with no emitter, so a tool cannot stream a multi-minute run
into the conversation. One conversation list means one agent.
### Why the beat is a subagent rather than prose in one prompt
It started as a section appended to banking's prompt. That left per-beat
configuration homeless (model, effort and recursion limit are all
agent-level, and there was one agent), and made banking's
~21,000-character rulebook ride every one of the ~20 model calls the run
makes — re-sending rules about markdown tables while the agent read a
CSV.
Reached as a `CompiledSubAgent`, because a raw `SubAgent` spec has no
`subagents` field and this one needs its own: the per-merchant fan-out
is a headline of the beat, and a flat subagent could only research
serially. A probe confirmed nesting survives — the analyst's `task`
dispatches, the researchers' `search_merchant` calls and the final
report tool all reach `astream_events`.
## The console
One CLI window in the transcript carries the whole run: narration, `$
execute`, `search "…"`, `→ merchant-researcher: …`, results, indented by
depth so a ten-way fan-out reads as a fan-out. Subagent narration is
suppressed from the conversation so the console is the single place the
harness is visible; banking's own replies still render normally.
`shell/subagents/subagent-activity.tsx` subscribes to the agent's
**event stream**. Reading `agent.messages` — the first design — was
wrong twice: messages materialise at the `MESSAGES_SNAPSHOT` (two per
run), so the pane sat still for minutes then filled at once; and
persisted messages carry no `subagentRunId`, so the harness's narration
could not be told from banking's reply. The fold is pure and idempotent,
so the same code serves the live subscription and a replay of the
thread's stored events when a conversation is reopened.
Three heuristics were deleted along the way, each replaced by identity
the protocol already carried:
| heuristic | replaced by |
|---|---|
| console anchored on "the first tool call" | the run's first `task`
call, from **message order** (durable; the event-derived version
rendered one console *per delegation* on a restored thread — six,
measured) |
| `CONSOLE_TOOL_NAMES` allowlist | presence of `subagentRunId` |
| `disable_streaming` on the researchers | the canary's per-lane state |
The message filter suppresses **prose** and keeps **tool calls**: an
agent routinely narrates and calls a tool in one message, and returning
`null` for the whole message hid the report card — the run looked
perfect and ended with nothing to show.
## Canary stack, contained to this app
The subagent surface only exists on the canary line, and a released
`@ag-ui/client` ≤ 0.0.57 rejects `SUBAGENT_*` events **in the HTTP
transport before any middleware runs**, killing the stream. So the demo
leaves the root pnpm workspace and ships its own lockfile:
- `@copilotkit/* 1.68.3-canary.1786716392`
- `@ag-ui/* 0.0.59-canary.1786716392.0`
- `ag-ui-langgraph 0.0.43.dev1786716392` → `ag-ui-protocol
0.1.20.dev1786716392`
That keeps an unreleased protocol out of every other package in the
monorepo.
A 1.62.2-based canary was tried first and **could not compile the app**:
`workspace:*` is not a version, so the demo had no recorded lower bound
on the CopilotKit API it needs, and pointing it at 1.62.2 silently
rewound that API five minors. `OpenGenerativeUIActivityRenderer` — a
public `/v2` export since ~1.63 — was the first thing to break, and
there was no reason to think it was the only one. Rebasing the canary on
1.68.3 collapsed that whole class of risk.
## Verified against the running app
Every row is a measurement, not a claim.
| | |
|---|---|
| Registry shape | `/info`: `banking` = HttpAgent, other six =
BuiltInAgent |
| Sandboxed execution | 8 `execute` calls; agent writes and runs its own
Python |
| Parallel subagents | 6 `SUBAGENT_STARTED`/`FINISHED` pairs, 10
`search_merchant` |
| Canvas beat | `render_report` → a2ui middleware emits `activityType:
a2ui-surface` |
| Frontend tools | given two, picks `showTransactions`, emits no result,
emits **no prose** — the prompt's restraint rule surviving the port |
| Memory (Intelligence) | run is handed `recall_memory`, `save_memory`,
`forget_memory`, `knowledge_base_shell` |
| HITL round-trip | tool call out, answer back in, agent continues |
| Durable background run | client disconnected at 8s; run finished
unattended; thread replayed |
| Thread restore | 61 messages persisted incl. the report tool (an
earlier flat-subagent attempt collapsed to 4) |
| Report correctness | 14 rows, 9 researched, 6 filed with ids read out
of real 201 bodies, totals reconciling against their own rows |
| Run duration | 86s reported vs 98s wall clock |
## Correctness bugs found by running it
Each produced a confident, complete-looking wrong answer rather than an
error — the characteristic failure of a multi-minute agentic beat.
- **Beat 3d was dead, and it took the whole thread with it.** The bet
against this one was right. `ag_ui_langgraph` routes *every* attachment
to the model as an `image_url` block, documents included, so the Q2
invoice was rejected before the first token: `400 Invalid MIME type.
Only image types are supported`. The exception is raised inside the
model node, which kills the SSE stream — the runtime sees `RUN_ERROR:
terminated` with no cause and the browser renders **nothing**: no error
bubble, no failed message. And the crashed run is still checkpointed, so
every later message on that thread replays the rejected content and dies
the same way. One click on the pill killed the conversation permanently;
only restarting the service cleared it. Fixed here by a
`wrap_model_call` middleware that rewrites those blocks into LangChain
standard `file` blocks, and upstream in
[ag-ui#2476](https://github.com/ag-ui-protocol/ag-ui/pull/2476) (both
adapters, plus the return leg so a non-image attachment survives
`MESSAGES_SNAPSHOT`). The middleware is a stopgap with its deletion
condition written into its header — this service installs the adapter
from PyPI, so the upstream fix cannot reach it until published.
- **Totals did not match their own rows.** Every per-row amount matched
the CSV while the headline totals came back $1.00 and $0.20 high: the
model authored them instead of adding them. The card prints the total
directly above the rows it is the sum of. Fixed structurally — derived
in `submit_expense_report`, removed as parameters.
- **`amount` arrived as a string** on all 14 rows, silently defeating
`toLocaleString(…currency…)` so it printed `842.10` with no currency.
- **`merchantKind` non-answers.** With no search tool the model wrote a
bare `"unclear"`; with Tavily live it hedges in prose (`"unknown (likely
bookbindery/bookshop retail, but not established for this exact
merchant)"`), which an exact-match filter passed into a 60-character
label glued to the merchant name. Now rejected on leading token, hedging
language, and a 40-character cap.
- **The run clock reported 333s for a two-minute run.** It took the
oldest open stamp across the process because the tool could not name its
own run; model calls *after* the report re-stamped it and the leftover
became the next run's start. Now keyed per run via an injected
`ToolRuntime`.
- **`graph.with_config({"recursion_limit"})` is silently dropped** by
the AG-UI adapter. The agent completed the entire analysis, streamed
every argument of the final report, then died at LangGraph's default of
25 supersteps.
- **A sync `wrap_model_call` under `astream`** surfaces as a bare
`RUN_ERROR: terminated`, cause only in the service log.
- **`emit_raw_events` defaults to `True`**, piggybacking LangChain's
internal events onto the stream: 27,950,261 bytes → 374,086 with it off,
identical report. Matters because the thread *persists* those events for
replay.
- **`gpt-5.6-sol` + `reasoning_effort` + function tools 400s** on
`/v1/chat/completions`; needs `use_responses_api=True`. The first model
probe missed it by binding no tools — a model probe for an agent has to
bind one.
## Upstream findings (reported separately, not fixed here)
1. **`@copilotkit/runtime` drops `subagentRunId` when persisting
messages.** 2888 of 3026 stream events carry it; 0 of 53 persisted
messages do. Reproduced with Intelligence removed entirely, so it is the
runtime's message shape rather than the platform store — and
`@copilotkit/runtime`'s dist contains no occurrence of the field, while
`@ag-ui/core`, `ag-ui-protocol` and `@copilotkit/core` all model it. One
field threaded through would let the console rebuild from message
history and delete the event-replay seeding here.
2. **`copilotkit` 0.1.95 × `ag-ui-langgraph` 0.0.43** — the FastAPI
endpoint calls `agent.clone()` per request; 0.0.43's base `clone()`
hard-passes three kwargs the SDK subclass does not accept, so **every
request 500s**. Verified with a minimal repro on stock classes and by
reading published wheels (0.0.41/0.0.42 are fine — the window is 0.0.43
only). `sdk-python` requires `>=0.0.42` unbounded, so fresh installs
break, and `examples/showcases/deep-agents{,-finance-erp,-job-search}`
are one `uv lock --upgrade` away. Now open as #6592 (`**kwargs`
passthrough + 9 regression tests incl. a forward-compat guard); it needs
a `0.1.96` bump to reach PyPI before `main.py`'s
`BankingAGUIAgent.clone()` workaround can go — and note `agent/uv.lock`
pins `copilotkit 0.1.95`, so removing the workaround is a re-lock as
well as a delete. That branch also fixes 4 pre-existing failures in the
sdk-python suite from the same root cause — **`test_unit-python-sdk` may
currently be red on `main`; worth checking independently.**
3. **Both LangGraph AG-UI adapters send non-image attachments as
`image_url`**, so a PDF, audio clip or video is rejected on the block
kind — see beat 3d in *Correctness bugs*. Fixed in
[ag-ui#2476](https://github.com/ag-ui-protocol/ag-ui/pull/2476); carried
here as a middleware until that publishes.
Also verified and **not** a problem: pnpm 10.33.4 still applies
`package.json` `pnpm.overrides` despite warning that it ignores them —
the ~70 root overrides including the security pins are live, nothing was
silently unpinned. (Migration to `pnpm-workspace.yaml` is worth doing
anyway, since the installed pnpm is 11.21.0 where the field genuinely is
dropped; branch exists, byte-identical lockfile.)
## What is not done
1. **Docs are stale.** `CLAUDE.md` still says `AgentRegistration` is `{
createAgent: () => BuiltInAgent }` and that the route "builds one
`BuiltInAgent` per registered skin"; the reskin skill says the same in
three places, and `templates.md` scaffolds it. The real type is `() =>
AbstractAgent`. This app has a standing rule that every change answers
whether the skill went stale. It did — the skill's launcher step is
fixed, the `BuiltInAgent` claims are not.
3. **Pre-release dependencies**, now on both halves: the JS canaries in
this app's own `pnpm-lock.yaml`, and `ag-ui-langgraph` /
`ag-ui-protocol` pinned `==` to the matching `.dev` canaries in
`agent/pyproject.toml`. Both move to stable when the subagent work
ships.
4. `pnpm test:e2e` has not been run.
### Closed since this PR was opened
- **The demo has its own CI job** —
`.github/workflows/test_reskinnable-demo.yml`. Nx discovers projects
*through* the pnpm workspace (no `workspaceLayout` in `nx.json`), so
leaving it also left the repo-wide `nx run-many -t build`
(`static_compat.yml`), `-t check-types` (`static_quality.yml`) and `-t
test` (`test_unit.yml`) sweeps, and both static workflows carry
`paths-ignore: ["examples/**"]` besides. One job runs the four gates
(lint, typecheck, 2460 unit tests, build); a second syncs
`agent/uv.lock` with `--frozen` and asserts the subagent *capability*
rather than a version string, so it keeps meaning something after the
pin moves.
The build gate is the one a developer can least run locally: `next
build` corrupts a concurrently running dev server's PostCSS/Turbopack
cache — measured, `globals.css` transforms to garbage, every route 500s,
and a dev-server restart does not clear it because the corruption is on
disk.
**Its first two runs failed, both usefully**, and both bugs were
pre-existing:
1. `pnpm/action-setup@v6.0.10` with `package_json_file:` pointed at this
app still installed the **root's** pnpm 10.33.4 rather than this app's
pinned 10.10.0 — the resolver that ignores `pnpm.overrides`, which is
the only place three of the five `@ag-ui/*` canaries are pinned. Caught
only because the job asserts `pnpm --version` against the pin; it would
otherwise have installed under the wrong resolver and stayed green. Now
activated through corepack, which reads the nearest `package.json`.
2. `pnpm install` in this directory **never installed this app**. pnpm
walks *up* for a workspace root, found the repo's, and installed all 70
monorepo projects (4645 packages) while leaving this directory with no
`node_modules` — the next command failing as `sh: 1: eslint: not found`.
So the app that ships its own lockfile was uninstallable by its own
documented instruction. Fixed by giving it its own
`pnpm-workspace.yaml`; `ignore-workspace=true` in an `.npmrc` does
**not** work (CLI-only in pnpm 10.10, measured).
- **The canary `overrides` now live in `pnpm-workspace.yaml`**, their
supported home, as a side effect of that fix. They no longer depend on
`packageManager: pnpm@10.10.0` being the version that still reads
`package.json` — which matters because `@ag-ui/core`, `@ag-ui/encoder`
and `@ag-ui/proto` are pinned nowhere else and a released `@ag-ui/core`
rejects `SUBAGENT_*` in the HTTP transport.
- **All ten beats walked in a browser**, including 3d — which was
broken, for the reason in *Correctness bugs* above.
- **`./run-demo.sh` starts a complete demo.** It now brings up the
Python agent between the compose wait and `pnpm dev`, guarded on
`/health` so a re-run reuses a live one, and dies with `(cd agent && uv
sync)` when the venv is missing. `./stop-demo.sh` had the mirror gap and
now stops it too — that one mattered more than it looks: the start
script *reuses* a live :8124, so an orphan surviving teardown is
silently adopted by the next cold start, serving whatever code it was
launched with.
- **The agent's Python deps are pinned and locked.** They were not, and
the JS half was — so a colleague's fresh `uv sync` resolved the
*release* `ag-ui-langgraph 0.0.43`, which does not accept
`emit_subagent_events` and ships no subagent symbols. That is this PR's
headline feature, and it would have failed silently: `main.py` sets the
flag as an attribute, so the assignment succeeds against an object
nobody reads and the service starts clean. Now `==` pins plus a
committed `agent/uv.lock`, verified by building a venv from only the
tracked files and replaying a captured Q2-with-PDF payload through it.
- **The README told people to install from the repo root** "as a
workspace package". It is deliberately not one, so a root install did
nothing for this app — and installing in the right place did nothing
either, until the `pnpm-workspace.yaml` above.
## Relationship to #6501
Based on `b94e4bfb5d`, the last commit before Codex appears, which is
**not on `main`**. So this PR carries **9 commits: the 4 here plus the 5
foundation commits it shares with #6501** — harness types + fixture,
prompt/workspace, the OFFSITE-to-fixture invariant guard, the
summary-shape/filing-contract fix, and `POST /transactions`. Whichever
merges first shrinks the other. #6501 and #6565 are deliberately
untouched.
## A demo-design question, not a bug
The fixture's merchant names are invented, so `Cardinal & Ash` — the
prompt's own worked example of *"could be a restaurant or a law firm,
find out"* — cannot be resolved by real web search and stays `unclear`
alongside `Bluebonnet Provisions`. The beat's headline claim is that the
agent researches every merchant; real merchant names in the CSV would
make that land harder.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
https://claude.ai/code/session_01JRubZT6AS6LCGkcE2KzcfA
`pnpm install` in this directory did not install this app. Leaving the
root workspace's member list was only half of it: pnpm walks UP from the
cwd for a workspace root, found the repo's, and installed THAT — measured
in CI (run 32398378642), "Scope: all 70 workspace projects", 4645 packages
resolved for the monorepo, and this app left with no node_modules. The
next command then failed as `sh: 1: eslint: not found`, which reads as a
broken toolchain rather than an install that went elsewhere.
A `pnpm-workspace.yaml` here stops the walk. `ignore-workspace=true` in an
`.npmrc` does not — it is CLI-only in pnpm 10.10 (measured: the Scope line
was unchanged), so the alternative would have been a flag every human and
job had to remember.
The five canary `overrides` move into that file, their supported home.
They stay duplicated in `package.json`'s `pnpm` field for now, which is
read only because this app pins `packageManager: pnpm@10.10.0` — three of
the five (`@ag-ui/core`, `@ag-ui/encoder`, `@ag-ui/proto`) are pinned
nowhere else, and a released `@ag-ui/core` rejects SUBAGENT_* events in the
HTTP transport, so a packageManager bump would have silently killed the
harness console.
Verified read-only: `pnpm install --frozen-lockfile --lockfile-only` in
this directory resolves the single project, satisfies the committed
lockfile, and leaves it byte-identical.
A colleague cloning this branch could not reproduce the demo. The JS half
is pinned exactly — this app left the root pnpm workspace and ships its
own `pnpm-lock.yaml` with `@ag-ui/client 0.0.59-canary.1786716392.0` — but
the Python half pinned nothing: `ag-ui-langgraph>=0.0.43` and no
`uv.lock`, so a fresh `uv sync` resolved the RELEASE.
Measured, on the release that `>=0.0.43` actually selects:
ag-ui-langgraph==0.0.43
emit_subagent_events accepted by LangGraphAgent.__init__: False
subagent symbols in ag_ui.core: NONE
That is this branch's headline feature — the streaming CLI console — and
it would have failed SILENTLY. `main.py` sets `emit_subagent_events` as an
attribute (copilotkit's subclass takes only four kwargs), so on a release
without the feature the assignment succeeds, lands on an object nobody
reads, and the service starts clean. No `subagentRunId` reaches the
browser, the console cannot separate the harness's work from the parent's,
and a reopened thread collapses a multi-minute run to one tool message.
Every gate stays green.
So `ag-ui-langgraph` and `ag-ui-protocol` are now `==` pins, and
`agent/uv.lock` is committed. `ag-ui-protocol` is pinned as a DIRECT
dependency although nothing imports it by name: it carries the SUBAGENT_*
event types and the adapter asks only for `>=0.1.15`, so left transitive
it resolves the release and undoes the other pin.
README: the quick start said `pnpm install # from the repo root — this is a
workspace package`. It is not one — it is absent from
`pnpm-workspace.yaml`, deliberately, so the canary line cannot leak into
the rest of the monorepo. A root install therefore installs nothing for
this app, which is a confusing first five minutes for anyone who reads it
and follows it.
Verified by cold start rather than by inspection: copied ONLY the files
git tracks (the five .py modules, pyproject.toml, the new uv.lock) into an
empty directory, ran `uv sync --frozen`, and got
ag-ui-langgraph 0.0.43.dev1786716392 with `emit_subagent_events accepted:
True` and the five SUBAGENT symbols present. Then booted that venv on a
spare port and replayed the browser's captured Q2-with-PDF payload
through it: RUN_FINISHED, with createReport carrying the invoice's real
line items. The live stack was not touched.
Reskin skill: checked, no impact. Its install/verify step is `pnpm dev`
inside this app, which is correct either way; the root-vs-here distinction
is a README concern and the skill never mentions the workspace.
The teardown mirror of the previous commit. `./stop-demo.sh` stopped the
dev server, the docker stack and the native TEI — never :8124 — so
banking's Python agent survived every teardown.
That leftover is not merely litter. `run-demo.sh` health-checks :8124
before starting (so a re-run reuses a live agent instead of colliding on
the port), which means the next cold start silently ADOPTS the orphan and
serves whatever code it was launched with. Edit `agent/`, re-run the
script, observe no change, conclude the edit did nothing.
No `--keep-agent` flag to match `--keep-tei`: TEI has that flag because it
is slow to warm, and the agent boots in seconds, so keeping it would only
reintroduce the failure above.
Also corrects two things in the same breath:
* The Ctrl-C claim I got backwards one commit ago. MEASURED this time,
with the same shell construct the script uses: SIGINT reaches the
foreground process group, which the backgrounded children are still in,
but a NON-INTERACTIVE shell sets background jobs to ignore SIGINT
(POSIX) — so only the exec'd dev server dies (exit=-2) and the stack,
TEI and agent all survive. `nohup` is not what saves them; that covers
SIGHUP, a different signal. Both scripts and the README now say this.
* `ok "docker stack down${PURGE:+ (volumes removed)}"` printed "(volumes
removed)" on EVERY teardown, because the flag holds the string "0" when
unset and `:+` expands on non-empty. The action was always right
(`--volumes` is gated on `-eq 1`) — verified: the postgres/redis/minio
volumes are still there after a flagless run — but the line told anyone
reading it that their seeded data had just been deleted.
Verified with a full cycle through both scripts: `./stop-demo.sh
--keep-tei` reported the agent stopped and left the volumes in place, then
`./run-demo.sh` came back with `banking agent ready (200)` and `stack
healthy`. The idempotency guard was exercised against the live service and
reports "already up" rather than starting a second uvicorn.
Reskin skill: checked, no impact. It documents authoring a skin, not
running the stack; its one launcher line got the note it needed in the
previous commit.
`./run-demo.sh` brought up the embedder, the Intelligence stack and the
dev server, then handed over an app whose DEFAULT skin could not answer a
single message. Banking's agent is a Python service (`agent/`, :8124) and
it is not a compose service, so it had to be launched by hand — and
nothing said so: no line in the script, no line in the README, no line in
any markdown in this tree.
The failure mode is the expensive kind. Nothing errors at startup: the
stack comes up healthy, the app boots, the dashboard renders off the REST
ledger, every pill is present. Only sending a message fails, and the six
in-process skins keep working, so the obvious reading is "my machine is
fine, the demo is broken".
The script now starts the agent between the compose wait and `pnpm dev`,
guarded on `/health` so a re-run reuses a live one instead of colliding on
the port, and dies with `(cd agent && uv sync)` when the venv is missing —
the same shape as the native-TEI branch above it.
Also corrects the Ctrl-C line, which claimed Ctrl-C "stops only the dev
server". Measured: the docker stack survives, and the dev server, the
native TEI and now the agent all go down with the script.
README: the quick start said `pnpm dev` and described OSS mode as needing
only `OPENAI_API_KEY`. True for six skins, false for the default one. It
now starts the agent too and says why the whole agent lives out of
process.
Reskin skill: updated, one sentence in the Verification list's step 2. A
skin author runs `pnpm dev`, gets redirected from `/` to banking, sends a
test message to see if anything works, and gets silence — then debugs
their own registration. The skill now points them at `/<their-id>` or
`./run-demo.sh`.
Verified by doing it: stopped the dev server and the agent, re-ran the
script, and it reported `banking agent ready (200)` and `stack healthy`
without touching the warm TEI. Then walked beat 3d in the browser — the
Q2 pill filed a report citing "the Meridian Creative Agency invoice from
page 1", i.e. the model read the attached PDF.
Beat 3d was dead on banking: clicking the Q2 pill staged the invoice,
sent the message, and then nothing happened at all — no report, no error,
no failed message in the transcript.
`ag_ui_langgraph` hands every attachment to the model as an `image_url`
block, documents included, so the PDF was rejected before the first
token:
openai.BadRequestError: 400 - Invalid MIME type. Only image types are
supported. (code: invalid_image_format)
The exception is raised inside the model node, which kills the SSE
stream. The runtime sees `RUN_ERROR: terminated` with no cause and the
browser renders nothing. Worse, the crashed run is still checkpointed, so
every LATER message on that thread replays the rejected content and dies
the same way: one click on the pill killed the whole conversation, and
only restarting this service cleared it (`MemorySaver` is in-process).
`_repair_document_attachments` rewrites those blocks into LangChain
standard `file` blocks before the model call. It walks every message, not
just the newest, because the offending content comes back from the
checkpoint on later turns.
STOPGAP, not a design. The real fix is open upstream as
ag-ui-protocol/ag-ui#2476 (both adapters, plus the return leg so a
non-image attachment survives MESSAGES_SNAPSHOT). This service installs
the adapter from PyPI, so that fix cannot reach this venv until it is
published; the middleware's header says when to delete it.
No test: `agent/` has no python test runner, and standing one up for code
whose deletion is already scheduled is the wrong trade. The durable
tests — PDF, audio, video, filename, round-trip, legacy binary — ship
with the upstream PR instead. Verified here by replaying the browser's
real captured run payload against the service: 400 before, RUN_FINISHED
with `createReport` carrying the invoice's line items after.
Reskin skill: checked, no impact. Its beat-3d guidance is entirely the
CLIENT half (staging into the composer, the `AttachmentFailureCause`
union, do not copy `@/shell/attach`), and it names no model-side
conversion. A skin authored from the skill gets a `BuiltInAgent`, whose
converter already maps documents onto file parts — banking is the only
skin whose agent is a LangGraph service, so this failure is unreachable
from the skill's path.
Gates: lint clean, typecheck clean, 2460 unit tests pass. `pnpm build`
deliberately not run — the diff touches no TypeScript, and `next build`
would clobber the `.next` state of the dev server currently serving the
demo.
Makes the offsite-expenses beat legible while it runs, and gives it its own
model, by taking the AG-UI subagent surface from the canary line.
## The expense analyst is now a real subagent
banking gpt-5.4, temp 0
│ banking's prompt; browser frontend tools; Intelligence memory tools
└─ expense-analyst gpt-5.6-sol, reasoning_effort=high
│ sandboxed shell, submit_expense_report
└─ merchant-researcher gpt-5.4, one per merchant, Tavily
Previously the beat was a section of banking's prompt, which left nowhere to put
per-beat configuration: model, effort and recursion limit are all agent-level and
there was one agent. It also meant banking's ~21,000-character rulebook rode
every one of the ~20 model calls the run makes, re-sending rules about markdown
tables while the agent read a CSV.
Reached as a `CompiledSubAgent` because a raw `SubAgent` spec has no `subagents`
field and this one needs its own — the per-merchant fan-out is a headline of the
beat, and a flat subagent could only research serially. Verified nesting
survives: a probe showed the analyst's `task` dispatches, the researchers'
`search_merchant` calls and the final report tool all reaching `astream_events`.
`gpt-5.6-sol` additionally needs `use_responses_api=True`: with function tools
and `reasoning_effort` it 400s on /v1/chat/completions. The first probe missed
that by asking the model a question with NO tools bound — a model probe for an
agent has to bind one.
## The console: one CLI window, streaming
`shell/subagents/subagent-activity.tsx` subscribes to the agent's event stream
and folds it into console lines. Reading `agent.messages` (the previous design)
was wrong twice over: messages materialise at the `MESSAGES_SNAPSHOT`, two per
run, so the pane sat still for minutes and then filled at once; and persisted
messages carry no `subagentRunId`, so the harness's narration could not be told
from banking's own reply.
The fold is pure and idempotent — every line keyed by the id of the thing that
produced it — so the same code serves the live subscription and a replay of the
thread's stored events when a conversation is reopened.
Three heuristics are deleted, each replaced by identity the protocol already
carried:
- the console's "first tool call" anchor -> the run's first `task` call, from
MESSAGE order (durable; the
event-derived version rendered
one console per delegation on a
restored thread — six, measured)
- `CONSOLE_TOOL_NAMES` suppression list -> `subagentRunId` presence
- `disable_streaming` on the researchers -> the canary's per-lane state
`shell/subagents/subagent-message-filter.tsx` keeps subagent narration out of the
conversation. It suppresses the PROSE and keeps the TOOL CALLS: an agent
routinely narrates and calls a tool in one message, and returning null for the
whole message hid the REPORT CARD — the run looked perfect and ended with nothing
to show. Shell-level and inert for a skin whose agent has no subagents.
## Canary stack, contained to this app
The subagent surface only exists on the canary line, and a released
`@ag-ui/client` <= 0.0.57 rejects `SUBAGENT_*` events in the HTTP transport
before any middleware runs, killing the stream. So the demo leaves the root pnpm
workspace and ships its own lockfile, pinning `@copilotkit/* 1.68.3-canary` and
`@ag-ui/* 0.0.59-canary` locally instead of imposing an unreleased protocol on
every package in the monorepo.
A 1.62.2-based canary was tried first and could not compile the app: it silently
rewound the CopilotKit API five minors under a demo written against 1.67.1, and
`OpenGenerativeUIActivityRenderer` (a public `/v2` export since ~1.63) was the
first thing to break. `workspace:*` is not a version, so the app had no recorded
lower bound on the API it needs.
KNOWN GAP, deliberately not fixed here: Nx discovers projects THROUGH the pnpm
workspace (there is no `workspaceLayout` in `nx.json`), so leaving it also
removes the demo from the repo-wide `nx run-many -t build` and `-t check-types`
sweeps. Verified — `nx show project deep-agents` and the other standalone
showcases return "Could not find project". No workflow names this demo, so it is
currently unbuilt and untype-checked in CI and needs its own job. Run the four
gates locally until that lands. Documented in `pnpm-workspace.yaml`.
## Fixes
- The run clock is keyed per run and read through an injected `ToolRuntime`
instead of taking the oldest open stamp across the process. Model calls AFTER
the report re-stamped the clock and that leftover became the next run's start:
a two-minute run reported 333s. Now 86s reported against 98s wall clock — the
gap is thread-naming and delegation, before the analyst's first model call,
which is what the tile claims to measure.
- `merchantKind` non-answers are rejected on the leading token, on hedging
language, and over 40 characters. With no search tool the model wrote a bare
"unclear"; with Tavily live it hedges in prose ("unknown (likely
bookbindery/bookshop retail, but not established for this exact merchant)"),
which an exact-match filter passed into a 60-character label glued to the
merchant name.
- `vitest` no longer externalises `@copilotkit/*`. Installing them from npm moved
`src/app/layout.tsx`'s stylesheet import under `node_modules/.pnpm/`, where
Node's ESM loader threw `Unknown file extension ".css"` and took out 16 suites
while naming a stylesheet nobody had touched.
- `agent/main.py` reads the demo's `.env` as well as its own, so `TAVILY_API_KEY`
works wherever an operator puts it. Two env files to keep in sync is a trap
whose failure mode is "the agent ignores a key that is plainly sitting in .env".
## Upstream finding (reported separately, not fixed here)
`@copilotkit/runtime` drops `subagentRunId` when persisting messages: 2888 of
3026 stream events carry it, 0 of 53 persisted messages do. Reproduced with
Intelligence removed entirely, so it is the runtime's message shape rather than
the platform store — and `@copilotkit/runtime`'s dist contains no occurrence of
the field at all, while `@ag-ui/core`, `ag-ui-protocol` and `@copilotkit/core`
all model it. One field threaded through would let the console rebuild from
message history and delete the event-replay seeding added here.
`CLAUDE.md`'s appended block is generated by `next dev`
(`next/dist/server/lib/generate-agent-files.js`) and committed per its own
instruction to keep the tree clean.
Gates: lint 0, typecheck 0, test:unit 2460 passed across 216 files, build 0.
Measured end to end in Intelligence mode: 14 rows, 9 merchants researched, 6
charges filed with ids read out of real 201 bodies, totals reconciling against
their own rows, 3220 events with 6 SUBAGENT_STARTED/FINISHED pairs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JRubZT6AS6LCGkcE2KzcfA
THE PILL. The offsite-expenses beat had no entry point in the UI. The agent was
wired, the service was up, all four gates were green, and every verification I
ran POSTed directly at the runtime — so the beat was fully working and completely
unreachable for anyone opening the app. It is the ninth and last pill, after the
AWS charge, because it is the only one that does not answer in seconds.
`suggestions.test.ts` now pins it. That guard is not decoration: this pill has no
type referencing it, no matcher and no route, so nothing else in the tree
notices if it is dropped or retitled. It checks REACHABILITY, which is the
property nothing else was holding.
MERCHANT KIND HARDENING. `merchantKind` is a KIND ("hotel", "pharmacy") that the
report card prints inline beside the merchant name, where it has room for about
two words. The previous filter rejected exact matches of "unclear"/"unknown",
which was enough when there was no search tool and the model wrote a bare
"unclear". With Tavily live it hedges in prose instead — measured:
"unknown (likely wellness-related business)" and "unknown (likely
bookbindery/bookshop retail, but not established for this exact merchant)". Both
sailed through and would have rendered as a 60-character label glued to the
merchant name.
Now rejects on the leading token, on hedging language anywhere, and on anything
over 40 characters. The length cap is the backstop: it encodes the actual
constraint rather than a list of phrasings already observed, so it catches the
next wording nobody predicted. A measured run after the change leaked none, with
a longest surviving kind of 23 chars.
ONE ENV FILE. `agent/main.py` now loads the demo's `.env` in addition to its own,
with `agent/.env` winning on conflict. The app and this service need the same
keys, and asking an operator to keep two files in sync is a trap whose failure
mode is "the agent ignores a key that is plainly sitting in .env" — which is
exactly what happened when TAVILY_API_KEY was added to the demo's `.env` and the
service, reading only its own, never saw it.
`.env.example` documents TAVILY_API_KEY and BANKING_AGENT_URL, including what a
keyless run actually looks like: the research subagents are told plainly that no
search happened and instructed to report "could not establish" rather than
guess, so the run completes with several rows marked `unclear` and a "merchants
researched" tile reading 0. Nothing errors, which is why it needs writing down.
Measured with Tavily live: 14 rows, 10 merchants researched, 10 parallel
subagents, 31s, totals reconciling against their own rows, 7 charges filed with
ids read out of real 201 bodies, zero non-answer kinds. Research also CHANGES a
decision rather than only labelling one — The Copper Room resolved to
`restaurant_lounge` and moved from `unclear` to `expensable`.
Gates: lint 0, typecheck 0, test:unit 2460 passed across 216 files, build 0.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JRubZT6AS6LCGkcE2KzcfA
Banking's `agent.ts` stops returning a `BuiltInAgent` and returns an `HttpAgent`
pointed at the Python deep agent instead. Six skins still run in-process;
banking is the one that does not.
WHY THE WHOLE AGENT, not just the offsite-expenses beat. Threads are scoped per
agent — `listThreads` takes `agentId` as a REQUIRED parameter, and measured
against the running app the two lists are disjoint (banking 46 threads,
banking-expenses 10). The v2 runtime has no handoff mechanism, and
`defineTool`'s `execute` takes `(args)` with no emitter, so a tool cannot stream
a multi-minute run. Under a second agent id a presenter could start the
analysis, switch threads, and have nothing to come back to. One conversation
list means one agent.
The expense task spec is appended to banking's own prompt as a CONDITIONAL job
rather than a second identity — it opens by saying it applies only when the user
hands over a statement, and to ignore it otherwise.
Ported to Python:
- `agent/prompt.py` — banking's system prompt, extracted from the TS template
literal programmatically and asserted equal, not paraphrased. 21,208 chars
carrying 56 distinct rules; a dropped rule silently breaks a beat that still
looks like it works.
- `agent/report.py` — `render_report` and the a2ui operations builder. Component
order, ids and the `columns` arithmetic match the TS builder; the unique
surface suffix uses uuid rather than a timestamp so two reports in the same
millisecond cannot collide.
Carried over from the TS `BuiltInAgent` because each had a reason written beside
it: the non-mini model (the teach-and-recall arc routes unreliably on mini) and
`temperature: 0` (tool routing must be deterministic).
Also sets `emit_raw_events = False`. It defaults to True, which piggybacks
LangChain's internal events onto the AG-UI stream: a measured run streamed
27,950,261 bytes, and the same run with it off streamed 374,086 — a 75x
reduction with an identical report. This matters more than it looks because the
thread PERSISTS those events for replay, and leaving a running thread and coming
back to it is the point of this beat. Nothing downstream reads RAW.
Verified against the running app, not reasoned about:
- `/info` reports banking as HttpAgent and the other six as BuiltInAgent.
- Canvas beat: `render_report` -> TOOL_CALL_RESULT -> the a2ui middleware emits
`activityType: a2ui-surface` with createSurface + updateComponents.
- Frontend-tool beat: given showTransactions + showPendingApprovals the agent
picks showTransactions, emits no result (the client executes it) and emits no
prose — the prompt's "the rendered list is the single source of truth" rule
surviving the port.
- Intelligence mode: the run is handed the browser's tool AND
recall_memory / save_memory / forget_memory / copilotkit_knowledge_base_shell.
- Durability: a run whose client disconnected at 8s finished unattended and left
a replayable thread.
Removes `expenses-agent.ts` and the non-skin `banking-expenses` registry key;
banking's own id now serves both. `AgentRegistration.createAgent`'s comment is
rewritten — it described a second remote key that no longer exists — and now
gives a grep to derive the BuiltInAgent/HttpAgent split rather than asserting it.
Gates: lint 0, typecheck 0.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JRubZT6AS6LCGkcE2KzcfA
Replaces the Codex CLI subprocess with a Python LangChain deep agent
(`examples/showcases/reskinnable-demo/agent/`) reached over AG-UI: sandboxed
shell execution via deepagents' `LocalShellBackend`, parallel per-merchant
research subagents, no external harness binary, and our own API key.
Registered as `banking-expenses` in the server agent registry. That required
widening `AgentRegistration.createAgent` from `() => BuiltInAgent` to
`() => AbstractAgent` — the runtime's own `agents` option is
`Record<string, AbstractAgent>` and the v2 tree contains no
`instanceof BuiltInAgent` branch, so the demo's type was stricter than the
runtime it described.
Measured end to end against the real runtime: 58 `execute` calls, 7 concurrent
`task` subagents (peak concurrency 4/4 on a dedicated probe), RUN_FINISHED,
eight charges filed with ids read out of real 201 bodies, and 8221 replayable
AG-UI events persisted on the thread after the client disconnected 8s in.
Correctness fixes found by running it rather than reading it:
- `submit_expense_report` DERIVES the totals from the verdicts instead of
accepting them as arguments. A measured run had every per-row amount matching
the CSV while the headline totals came back $1.00 and $0.20 high — the model
authored them. The report card prints the total directly above the rows it is
the sum of, so tiles and rows must not be able to disagree.
- `amount` is coerced at the tool boundary; it arrived as a string on all 14
rows, which silently defeats the widget's currency formatting.
- `merchantKind` values that are non-answers ("unclear") are dropped rather than
rendered beside the merchant name as if they were findings.
- The recursion limit is set on the AGENT, not via `graph.with_config(...)`,
which the AG-UI adapter drops — the run completed the whole analysis and then
died at LangGraph's default of 25 supersteps.
- A `clone()` override works around `copilotkit` 0.1.95 being incompatible with
`ag-ui-langgraph` >= 0.0.42, whose base `clone()` (called per request) passes
kwargs the SDK subclass does not accept. Every request 500s without it.
Deletes the superseded Node-side harness scaffolding: `prompt.ts` (the task
specification now lives in the Python agent's system prompt), `workspace.ts`
(the sandbox owns the workspace) and their tests. `types.ts` stays — it is the
shared vocabulary the report card and the tool renderer both import.
Also drops the three unused `@tanstack/*` dependencies.
Reskin-skill impact: checked. `.claude/skills/reskin/` documents the `Skin`
contract, and no field of it changed — `banking-expenses` is NOT a skin and must
never look like one. It is absent from `registry.ts`, `skinIds`,
`skinIdentities` and `LINTED_SKIN_IDS`, and has no route, theme or identity. The
one documented thing that did change is `AgentRegistration.createAgent`'s type,
which SKILL.md describes in its registration section; that is updated in the
follow-up commit that makes banking itself a deep agent.
Gates: lint 0, typecheck 0, test:unit 2458 passed across 217 files, build 0.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JRubZT6AS6LCGkcE2KzcfA
Three names for one value were live in CopilotKit's own documentation, and
following the wrong one with a CLI-provisioned project yields an undefined
key:
- `INTELLIGENCE_API_KEY` — what `copilotkit project select` writes, used by
all 34 integration examples and the docs site.
- `COPILOTKIT_INTELLIGENCE_API_KEY` — the seven Channels package READMEs and
the packaged skills. Nothing ever read it.
- `COPILOTKIT_API_KEY` — the Slack and Teams examples, and the TSDoc on
`CopilotKitIntelligence` itself, which is what an IDE shows on hover.
`INTELLIGENCE_API_KEY` wins, because it is the name the CLI provisions and
changing it would break every scaffolded project in the wild.
`COPILOTKIT_INTELLIGENCE_API_KEY` is retired outright — no code read it.
`COPILOTKIT_API_KEY` stays readable as a deprecated alias in the two
examples that consume it, so an existing `.env` keeps working, and is
documented as deprecated everywhere it appears.
The skills reference also documented `organizationId`, sourced from a fourth
and fifth env name, as a `CopilotKitIntelligence` option. It is not one:
`CopilotKitIntelligenceConfig` has no such field, so the copy-pasteable
sample it appeared in would not compile. Removed from the samples, and the
prose that told readers to fetch a value for it corrected.
The Intelligence wiring itself was published only inside
`node_modules/@copilotkit/runtime/skills/`, and the only docs pages showing
`CopilotKitIntelligence` were the two Channels frontends — so a developer on
the plain web path had no page to reach it from. Adds
`/premium/connect-your-runtime`, which covers the wiring, how to confirm the
credential is actually consumed, and the self-hosted two-URL rule.
`scripts/validate-intelligence-env-names.ts` keeps this from drifting back.
It runs unfiltered in CI on purpose: the two workflows that would otherwise
cover it filter paths, and static/quality ignores `examples/**` — exactly
where the deprecated alias lives.
Checked the reskin skill: no impact. This adds one banking-only REST route and
one store adder; it touches no Skin contract field, no shell file, no lint rule
and no gate the skill names.
Jerel John Velarde
