copilotkit 4.11.0 → 4.13.0

package/onboarding/prompts/research/gather.md

@@ -14,20 +14,40 @@ notification. Where your own harness hands you the report from the dispatch itse
 you already hold the result. Either way there is nothing to poll.
 
 After you dispatch, do the work that does not depend on the result -- dispatch the other
-subagent, run the preflight below -- and then stop and wait. Wherever a prompt in this graph
-says to wait for a subagent to finish, that is what it means.
+subagent, run the preflight the next prompt gives you -- and then stop and wait. Wherever a
+prompt in this graph says to wait for a subagent to finish, that is what it means.
 
 Do not sleep to pass the time. Not `sleep`, not `/bin/sleep`, not a timer under another
 name, and not a loop that re-checks whether a result has arrived. Sleeping tells you nothing
-that waiting for the result does not, and it spends wall clock, which is one of the things
-this journey is measured on. One recorded run spent most of two hours asleep between
-dispatches that had all returned in seconds.
+that waiting for the result does not, and it spends wall clock.
+
+## When a subagent cannot work
+
+This rule also covers every subagent this run spawns, here and in every later prompt.
+
+A subagent that returns no usable result has failed in your harness, not in this graph.
+Delegation buys parallelism and a clean context, not capability: every assignment in this
+run is work you can do yourself, slower. So run that assignment yourself and carry on,
+rather than treating it as the end of the run.
+
+Both research subagents failing means this harness has no working subagent at all, which
+is worth recording once:
+
+```text
+npx --yes copilotkit@4.13.0 onboard checkpoint --phase delegation-unavailable
+```
+
+Then say once, in your own words, that this environment has no working subagents, so you
+will carry out each step yourself and the run will take longer than usual.
+
+For the rest of this run, wherever a prompt tells you to spawn a subagent, read that
+`subagent/` brief yourself and follow it in place of spawning one.
 
 Before you ask the developer any setup question, finish every read-only investigation and
 preflight check in this section.
 
 Prepare two research assignments. Give each research subagent one assignment. Tell it to run
-`npx --yes copilotkit@4.11.0 onboard read subagent/inspect-repository` first and follow the
+`npx --yes copilotkit@4.13.0 onboard read subagent/inspect-repository` first and follow the
 prompt it returns. If that read fails because the subagent cannot use the shell, stop that
 subagent. Run the same command yourself, then spawn a fresh subagent with the returned prompt
 and the same handoff. Require only its assigned packet.
@@ -38,98 +58,20 @@ Start both research subagents in parallel:
 2. Spawn a second research subagent. Assign it the environment evidence packet.
 3. Start both research subagents. Then continue.
 
-Continue to the surface-control preflight.
-
-Prove whether your coding-agent environment has browser or device control. Do not assume it
-either way, and do not report what you expect to be true: a run that guesses here records a
-capability every later step then trusts. Use a browser or device tool already
-configured for the coding agent you are running as, the same way a later step uses the
-CopilotKit documentation server. With a browser tool, open one inert page such as
-`about:blank` and read its title. With a device tool and no browser, list the booted devices
-in one command: `adb devices -l`. Record the outcome of that attempt: `available` when the
-tool answered, `unavailable` when there was none to call or the page did not open. Use
-the control that matches the selected surface later.
-
-Where the browser probe answered, register nothing. The harness came equipped and the run
-owes it no setup.
-
-Where no browser tool answered, register one for the coding agent you are running as, then
-run the same probe again and record what the second attempt did. Register this server:
+Report that both subagents started:
 
 ```text
-npx --yes @playwright/mcp@latest --browser chrome --isolated --output-dir <project>/.copilotkit/proof/browser
-```
-
-Replace `<project>` with the absolute path of the target project directory. The server is
-registered against the coding agent rather than against a directory, so a relative path here
-resolves wherever that server happens to start.
-
-`--browser chrome` drives the Google Chrome the developer already has, so this downloads no
-browser. `--isolated` keeps the profile in memory, so it never touches their own Chrome
-profile. `--output-dir` is what keeps the snapshots and screenshots out of the developer's
-repository root: without it this server writes them to `.playwright-mcp/` beside their code,
-which a project that never asked for a browser has no reason to carry, and which this setup's
-own evidence rule already has a place for. Register it the way your own harness registers a
-server, which is the mechanism a later step uses for the CopilotKit documentation server.
-
-Tell the developer in one line what you registered, that it drives their installed Chrome,
-and that it lives in this coding agent's configuration rather than in their repository. Do
-not ask them to approve it, and do not ask a second question about it.
-
-Do not add a browser or device driver to the project. A driver added there is a
-devDependency and a browser download in the diff of a repository that never asked for one,
-which is a different thing from a server registered against the coding agent.
-
-Where the second probe still does not answer -- no Chrome to drive, no network, or a
-harness that cannot register a server -- record `unavailable` and carry it. Do not keep
-trying, and do not ask the developer about this tool limit.
-
-Registering a server does not boot a device. Where the device probe found none, that is the
-whole finding, and a browser is not a substitute for a device.
-
-Wait for both research subagents to finish.
-
-## Settle the port
-
-The findings name the ports this project declares and the ports already held on this
-machine. Pick one free port for the frontend now, and use that number for the rest of the
-run: in the plan, in the implementation, and in the proof. Record it with the findings you
-hand to every later subagent.
-
-Do not let a later step pick again. A run that decides the port three times produces three
-numbers, and `verify` reads none of them: with no recorded URL it assumes port 3000, and
-whatever holds that port answers. An unrelated checkout on this machine then reports as
-this project's proven wiring.
-
-Project selection is where the settled port is written down, through `--runtime-url`.
-
-Then report that the research came back:
-
-```text
-npx --yes copilotkit@4.11.0 onboard checkpoint --phase research-returned
+npx --yes copilotkit@4.13.0 onboard checkpoint --phase research-dispatched
 ```
 
 A refused checkpoint prints its reason and leaves onboarding unaffected. It is not a
 failed step.
 
-Continue only if both research results start with `Status: passed`. For `Status: failed`,
-retry only that packet with its failed items, up to three attempts. For `Status: blocked`,
-or a third failed result, use the stop route at the end of this prompt.
-
-Merge both evidence packets by item. On a conflict, spawn one fresh read-only verifier with
-the item, both cited findings, and the research limits. Require its result to start with
-`Status: passed`, `Status: failed`, or `Status: blocked`. Use only a passed cited result. Do
-not inspect the project to settle the conflict yourself. Use the stop route for a non-pass
-verifier result.
+The research is under way. Continue to the surface-control preflight while it runs:
 
-Match environment evidence to the target app directory from the project packet. Require one
-target app directory and one matching environment row. If either packet gives no match or
-more than one match, send each research worker a focused directory check. Continue only when
-both workers return the same one target app directory. Both results must start with
-`Status: passed`. Otherwise, use the stop route.
-
-When both research results are merged, run
-`npx --yes copilotkit@4.11.0 onboard read research/route`.
+```text
+npx --yes copilotkit@4.13.0 onboard read research/preflight
+```
 
 If inspection stops onboarding, run
-`npx --yes copilotkit@4.11.0 onboard read stopped/run-failed`.
+`npx --yes copilotkit@4.13.0 onboard read stopped/run-failed`.

package/onboarding/prompts/research/merge.md

@@ -0,0 +1,60 @@
+# Merge the research and settle the port
+
+The surface is settled. This phase collects both research packets, picks the port the rest
+of the run uses, and merges the findings into one answer.
+
+Wait for both research subagents to finish.
+
+## Settle the port
+
+The findings name the ports this project declares and the ports already held on this
+machine. Pick one free port for the frontend now, and use that number for the rest of the
+run: in the plan, in the implementation, and in the proof. Record it with the findings you
+hand to every later subagent.
+
+Do not let a later step pick again. A run that decides the port three times produces three
+numbers, and `verify` reads none of them: with no recorded URL it assumes port 3000, and
+whatever holds that port answers.
+
+Project selection is where the settled port is written down, through `--runtime-url`.
+
+Then report that the research came back:
+
+```text
+npx --yes copilotkit@4.13.0 onboard checkpoint --phase research-returned
+```
+
+A refused checkpoint prints its reason and leaves onboarding unaffected. It is not a
+failed step.
+
+Continue only if both research results start with `Status: passed`. For `Status: failed`,
+retry only that packet with its failed items, up to three attempts. For `Status: blocked`,
+or a third failed result, use the stop route at the end of this prompt.
+
+Merge both evidence packets by item. On a conflict, spawn one fresh read-only verifier with
+the item, both cited findings, and the research limits. Require its result to start with
+`Status: passed`, `Status: failed`, or `Status: blocked`. Use only a passed cited result. Do
+not inspect the project to settle the conflict yourself. Use the stop route for a non-pass
+verifier result.
+
+A target project directory that holds no project has no app directory to key on, and
+that is a merged result rather than a failed merge. Record that this run has no target app
+directory and no environment row, then take the read below. Do not send a focused directory
+check, and do not use the stop route: neither worker can find a directory a developer has
+not created yet, and the route below is where such a project picks its starter.
+
+A directory holding only a Git repository, a coding agent's own configuration, editor
+settings or scratch notes holds no project. `onboard inspect` reports that as
+`appDiscovery.reason` of `no-project`.
+
+For a target project that does hold an app, match environment evidence to the target app
+directory from the project packet. Require one target app directory and one matching
+environment row. If either packet gives no match or more than one match, send each research
+worker a focused directory check. Continue only when both workers return the same one target
+app directory. Both results must start with `Status: passed`. Otherwise, use the stop route.
+
+When both research results are merged, run
+`npx --yes copilotkit@4.13.0 onboard read research/route`.
+
+If inspection stops onboarding, run
+`npx --yes copilotkit@4.13.0 onboard read stopped/run-failed`.

package/onboarding/prompts/research/preflight.md

@@ -0,0 +1,75 @@
+# Prove what this environment can drive
+
+The research subagents are running. This phase settles what your coding agent can drive, so
+every later step reads one recorded answer instead of guessing. Do not change application
+code here, and do not wait for the research packets yet.
+
+Prove whether your coding-agent environment has browser or device control. Do not assume it
+either way, and do not report what you expect to be true. Use a browser or device tool already
+configured for the coding agent you are running as, the same way a later step uses the
+CopilotKit documentation server. With a browser tool, open one inert page such as
+`about:blank` and read its title. With a device tool and no browser, list the booted devices
+in one command: `adb devices -l`. Record the outcome of that attempt: `available` when the
+tool answered, `unavailable` when there was none to call or the page did not open. Use
+the control that matches the selected surface later.
+
+Where the browser probe answered, register nothing. The harness came equipped.
+
+Where no browser tool answered, register one for the coding agent you are running as, then
+run the same probe again and record what the second attempt did.
+
+Tell the developer in one line what you are about to register, that it lives in this coding
+agent's configuration rather than in their repository, and that they can remove it again.
+Say that before you run the command. The run does not ask again.
+
+Register this server:
+
+```text
+npx --yes @playwright/mcp@latest --browser chrome --isolated --output-dir <project>/.copilotkit/proof/browser
+```
+
+Replace `<project>` with the absolute path of the target project directory. The server is
+registered against the coding agent rather than against a directory, so a relative path here
+resolves wherever that server happens to start.
+
+`--browser chrome` drives an installed Google Chrome. Where this machine has none, use
+`--browser chromium` instead and expect Playwright to download a browser build. Tell the
+developer before that download starts. If the server then reports no build, run
+`npx playwright install chromium`, and if it still cannot find one, run
+`npx @playwright/mcp install-browser chrome-for-testing`. `--isolated` keeps the profile in
+memory, so it never touches their own Chrome profile. `--output-dir` keeps the snapshots and
+screenshots out of the developer's repository root: without it this server writes them to
+`.playwright-mcp/` beside their code, which a project that never asked for a browser has no
+reason to carry. Register it the way your own harness registers a server, which is the
+mechanism a later step uses for the CopilotKit documentation server.
+
+Some coding agents, Claude Code among them, load a newly registered MCP server only at the
+next session start. Tell the developer in one line to expect one restart, then restart and
+re-bind with `npx --yes copilotkit@4.13.0 onboard start --run <onboarding_run_id>`. That
+restart is a step here, not an error.
+
+Do not add a browser or device driver to the project. A driver added there is a
+devDependency and a browser download in the diff of a repository that never asked for one,
+which is a different thing from a server registered against the coding agent.
+
+Where the second probe still does not answer -- no browser to drive, no network, or a
+harness that cannot register a server -- record `unavailable` and carry it. Do not keep
+trying. Say so in one line. Nothing here is for the developer to decide.
+
+Registering a server does not boot a device. Where the device probe found none, that is the
+whole finding.
+
+Report that the probe settled, whichever way it came out:
+
+```text
+npx --yes copilotkit@4.13.0 onboard checkpoint --phase surface-probed
+```
+
+Then merge the research:
+
+```text
+npx --yes copilotkit@4.13.0 onboard read research/merge
+```
+
+If inspection stops onboarding, run
+`npx --yes copilotkit@4.13.0 onboard read stopped/run-failed`.

package/onboarding/prompts/research/route.md

@@ -9,7 +9,7 @@ sends the run down one path.
 Before you route on, run this from the target app directory:
 
 ```text
-npx --yes copilotkit@4.11.0 onboard protect
+npx --yes copilotkit@4.13.0 onboard protect
 ```
 
 It reads the working tree itself, records every changed or untracked path with a digest,
@@ -28,6 +28,30 @@ graph exists to create is never stopped by them.
 If the result starts with `Status: blocked`, report the printed reason and use the stop
 route at the end of this prompt.
 
+## Route a named Channel before the findings
+
+The prompt the developer copied names the docs page it came from. A Slack or Teams docs
+page names the surface this run is for. So does a developer who has already said Slack or
+Microsoft Teams. Either one settles the frontend.
+
+Record that docs page and give it to every later node and subagent. The copied prompt is
+the only carrier of the surface the developer wants, and no flag states it.
+
+On either, run this before you read the three findings below, and without asking the
+purpose question:
+
+```text
+npx --yes copilotkit@4.13.0 onboard read feature/channels/start
+```
+
+Name the provider to that node so it does not ask again. A Slack page names Slack. A Teams
+page names Teams.
+
+Take this route whatever the three findings prove. An agent, a frontend and a CopilotKit
+integration are a valid start for a Channel, not a reason to convert the application: the
+run keeps that code as it is and wires a managed Channel onto it. Do not send a Channel
+request to the conversion route.
+
 ## What the merged findings must contain
 
 If a repository file exists, each finding must cite it. For an empty project, the subagents
@@ -37,6 +61,27 @@ The findings must cover what the project is for, the agent, frontend, CopilotKit
 authentication, credential names, and validation path. Do not ask the developer for facts
 that the repository answers.
 
+## What the CLI proposed for the framework and frontend
+
+`onboard inspect` matched the project's manifests against the framework and frontend sets
+this graph serves, and its `detection` block carries the result with the dependency and
+directory behind each answer. Read it as a proposal.
+
+- `match` names the one slug that fits. Confirm it against the findings before you use it.
+- `matched` holds more than one slug where more than one fits. `langgraph-python` and
+  `langgraph-fastapi` are two ways of serving one framework and no dependency tells them
+  apart, so pick by how the agent is actually served.
+- `match: null` with an empty `matched` means the CLI placed nothing. This is never a
+  stop, and never a reason to route out on its own. It says only that no dependency
+  matched. Whether the project can be onboarded is your judgment, not its.
+- `unmatched` names the dependencies it did not place. Use them as evidence for what the
+  project runs.
+
+Detection proposes and you dispose. Keep the documentation-gap exception exactly as it is:
+an agent framework outside this set that speaks AG-UI is still yours to recognize and
+route to best effort. Where your reading disagrees with the proposal, record your own
+answer and say that the two disagreed.
+
 ## Route on the merged findings
 
 Read the route off the merged findings. Three of them decide it, and you already hold all
@@ -51,7 +96,7 @@ settle these three from your own reading of the project. Each one comes from the
 packets or it is not proved.
 
 If all three are proved, prove the live starting state before any project file changes. Run
-`npx --yes copilotkit@4.11.0 onboard read proof/oss-baseline`.
+`npx --yes copilotkit@4.13.0 onboard read proof/oss-baseline`.
 
 Route there before you ask the developer anything else. The questions after this prompt
 select a framework and a frontend that the findings already name, so a developer who
@@ -64,7 +109,7 @@ developer nor the repository findings prove what the project is for, ask one gui
 question about the user outcome. This asks what the developer wants to build before you
 select a framework. Give two or three short examples and offer a minimal starter. Record
 the answer and give it to each later subagent. Then run
-`npx --yes copilotkit@4.11.0 onboard read credentials/plan`.
+`npx --yes copilotkit@4.13.0 onboard read credentials/plan`.
 
 Do not ask that question on the route above. A project carrying all three states its
 purpose in the application it already serves.
@@ -76,5 +121,5 @@ A purpose question here names a domain before that choice.
 Take the same read named above without asking.
 
 If authentication or inspection stops onboarding, run
-`npx --yes copilotkit@4.11.0 onboard read stopped/run-failed`. Neither says anything
+`npx --yes copilotkit@4.13.0 onboard read stopped/run-failed`. Neither says anything
 about whether this project's stack is supported, which is not yet known at this point.

package/onboarding/prompts/starter/clone.md

@@ -46,7 +46,7 @@ the derived name.
 Only if the developer asks for an existing project, or asks to see the projects they have,
 read the choices:
 
-`npx --yes copilotkit@4.11.0 project list --json`
+`npx --yes copilotkit@4.13.0 project list --json`
 
 Then ask which one to use. Do not order the projects by creation time. If the developer
 already gave this answer, do not ask again. Do not read a secret value. Do not show or
@@ -56,7 +56,7 @@ Run the command from the parent directory. Do not inspect another entry in the p
 directory. Replace each placeholder with the recorded value. Do not run a placeholder as
 a shell argument.
 
-`npx --yes copilotkit@4.11.0 init --name <project-name> --framework <framework-id> --channel none --no-banner --create <name> --install`
+`npx --yes copilotkit@4.13.0 init --name <project-name> --framework <framework-id> --channel none --no-banner --create <name> --install`
 
 Pass the confirmed name to both `--name` and `--create`: the app directory and its
 Intelligence project take the same name here. If the developer names an existing project,
@@ -72,7 +72,7 @@ account. The command does not need terminal input.
 Report the clone before you inspect anything:
 
 ```text
-npx --yes copilotkit@4.11.0 onboard checkpoint --phase starter-cloned
+npx --yes copilotkit@4.13.0 onboard checkpoint --phase starter-cloned
 ```
 
 This is its own step, not an aside. A run that clones and then goes quiet is
@@ -82,10 +82,10 @@ separates them.
 Do not rebuild the starter by hand. Then inspect only the generated
 paths inside the target directory. Record the files, install result, project connection,
 and validation commands. Then run
-`npx --yes copilotkit@4.11.0 onboard read proof/round-trip`.
+`npx --yes copilotkit@4.13.0 onboard read proof/round-trip`.
 
 If the command fails, report its exact error and do not claim that the starter is ready.
 Then run
-`npx --yes copilotkit@4.11.0 onboard read stopped/run-failed`. The starter is one this
+`npx --yes copilotkit@4.13.0 onboard read stopped/run-failed`. The starter is one this
 graph ships and the stack was chosen from its own supported list, so a command that
 returned an error is a run that broke, not a setup this release does not support.

package/onboarding/prompts/stopped/run-failed.md

@@ -23,13 +23,42 @@ Say what the developer has now. Name the processes still running and the files t
 changed, so they can carry on by hand or start again from a known state. A run that stops
 without saying what it left behind leaves the developer to discover it.
 
+## If one scoped fix can finish this run
+
+Do not decide this yourself. This run refused to widen its own scope, and that refusal
+stands. The developer is the one who can widen it.
+
+Name the one fix you propose, and say which file or process it touches. Ask the
+developer whether they approve that one fix. If they decline, or the question cannot be
+put to them, stop here and send the report below.
+
+If they approve it, make that one fix and nothing else. Then come back into this run:
+
+```text
+npx --yes copilotkit@4.13.0 onboard resume
+```
+
+Write the developer's approval to standard input, in one or two sentences: the fix they
+approved and what it touches. Send no secrets, source code, logs, or command output. The
+command refuses an approval that carries any of those, prints the reason, exits zero, and
+serves nothing. Reword it and send it again.
+
+The command prints the run id, the step it puts you back on, and that step's prompt.
+Follow that prompt. This is the same run, so do not run `onboard start` again, and keep
+the run id you already have.
+
+If the command refuses for any other reason, it says so and serves nothing. Report what it
+said and stop.
+
+## If no fix is approved
+
 Stop onboarding without making more repository changes.
 
 Send one short report. Run the friction command without another developer question. Do not
 ask the developer about telemetry: the command applies the setting they already have.
 
 ```text
-npx --yes copilotkit@4.11.0 onboard friction --phase stop --category <slug>
+npx --yes copilotkit@4.13.0 onboard friction --phase stop --category <slug>
 ```
 
 Write one or two sentences to standard input: the step you stopped at and what stopped

package/onboarding/prompts/subagent/create-plan.md

@@ -15,6 +15,23 @@ URLs you were given, not from another checkout on this machine.
 List the required credential variable names. Do not read or return credential values.
 Preserve each agent or frontend that already exists.
 
+## What the existing agent's behavior is
+
+The existing agent's behavior is four things: its system prompt and instructions, its
+tools and what those tools do, its model and provider configuration, and its memory or
+state handling.
+
+Wiring is not behavior. The CopilotKit SDK dependency, the middleware the agent is wrapped
+in, the runtime mount, the AG-UI connection, and a frontend that renders what the agent
+already returns are wiring. Plan them as ordinary steps.
+
+Every item on the behavior list is a change the developer has to approve. Where a step
+needs one, plan it as its own item, name the file and the exact change, and state that it
+changes how the agent answers everywhere, not only in CopilotKit. A system prompt the
+developer wrote is the agent's behavior rather than a setting this run tunes until a
+component renders. Where the plan cannot be written without such a change and you cannot
+state it that way, return `Status: blocked`.
+
 Plan the runtime to consume the Intelligence credential. The runtime takes an
 `intelligence` option holding a client built from the project key. A runtime given a
 `runner` option instead is the OSS runtime. It never reads the Intelligence key, and the
@@ -153,7 +170,7 @@ here is work the developer did not ask for.
 Plan the threads drawer itself: add it from the selected drawer page, where this frontend
 does not already render one. Where this journey's frontend framework ships no threads
 drawer -- React Native --, plan that the thread is proved by
-`npx --yes copilotkit@4.11.0 verify --round-trip`, which needs no browser. Do not plan a
+`npx --yes copilotkit@4.13.0 verify --round-trip`, which needs no browser. Do not plan a
 step that opens the managed Intelligence dashboard.
 
 ## Order the plan into steps
@@ -179,6 +196,12 @@ and one sentence saying what the step changes there and why no other file will d
 step in the plan. Return `Status: blocked` only when the plan needs a protected path you
 cannot justify in that sentence.
 
+Authorization covers changing a protected path, not removing it. Do not plan the deletion
+of a protected path, and do not plan a step that moves or renames one, which removes it
+under its captured name. If a step cannot be written any other way, return
+`Status: blocked` and name the path. The developer approves a plan that changes their
+files, and a plan that deletes one asks for something the audit cannot pass.
+
 Return the protected path list as part of the plan.
 
 Keep final validation and proof outside the implementation steps. The developer approves

package/onboarding/prompts/subagent/implement-and-validate.md

@@ -3,7 +3,8 @@
 Implement every step of the approved plan in plan order, and run the full validation list.
 Do not repeat the full repository inspection.
 
-Do not change a protected path.
+Do not change a protected path. Do not delete, move, or rename one either, whatever the
+plan says: a protected path that is gone fails the audit and no command clears that.
 Do not change a path that overlaps a protected path. Paths overlap when they are equal or
 either path is an ancestor directory on a path-segment boundary. `apps/a` overlaps
 `apps/a/src`, but not `apps/ab`. `app.ts` does not overlap `app.tsx`.
@@ -12,6 +13,13 @@ Change only the files and directories the approved plan names as changeable. If
 requires a file the plan does not name, stop and return the blocker. Do not run a
 repository-wide formatter.
 
+All implementation work happens in the run root's own working tree: the target app
+directory the main coding agent named. Never work in a separate worktree, a branch
+checkout, or a copy of the project. If any of your changes are already somewhere else,
+bring them into the run root before you report, and name them under Files changed. The
+completion step verifies that the changed files exist in this project, so work left in
+another tree ends the run with nothing in the developer's hands.
+
 Before validation, read the protected path list. Inspect the current changed and untracked
 paths. Keep protected paths out of the run path set. Compare every changed path with the
 paths the plan named. Here, a changed path means one in the run path set. If a changed path
@@ -25,6 +33,12 @@ Add only the props, options, and imports that appear in the fetched documentatio
 remembered API from an earlier CopilotKit version will fail type checking against this
 release. Do not add details that the fetched documentation omits.
 
+A symbol the installed package marks `@deprecated` in its type definitions is not a valid
+choice, whatever documentation found elsewhere shows. Read the installed type definitions
+for the symbol you are about to write. Where the selected documentation page itself shows
+a deprecated symbol, name that page under Blockers as `docs-wrong` friction and use the
+current symbol the package's own type definitions or changelog names.
+
 The documentation supplies the wiring. The project supplies the application. Use that
 wiring in the project's domain. Do not copy an example's agent name, tools, or data.
 
@@ -33,6 +47,23 @@ connection. Do not read, show, store, or return secrets. Do not read a file outs
 project directory: not for a credential, and not for an API question the fetched
 documentation answers. Ask the main coding agent for a missing credential.
 
+## What the existing agent's behavior is
+
+The existing agent's behavior is four things: its system prompt and instructions, its
+tools and what those tools do, its model and provider configuration, and its memory or
+state handling. Write none of them unless the approved plan named that exact change as its
+own item.
+
+Wiring is not behavior. Add the CopilotKit SDK dependency, wrap the agent in the
+middleware, mount the runtime, connect AG-UI, and render what the agent already returns.
+Those are the parts this run adds.
+
+A change on the behavior list changes how the agent answers everywhere, not only in
+CopilotKit, so the developer approves it before it is written. Where a step cannot be
+built without one the plan did not name, return `Status: blocked` and name the file, the
+change, and the check that asked for it. Do not edit the agent's prompt to make a
+component render.
+
 An existing `README.md` is the developer's, not this run's. Leave it exactly as you found
 it: not the title, not a section, not a line. Where this run has documentation to write,
 write it to a new file and name that file when you report the result. On an empty project