[feat] Agents name their own sessions and themselves - #5903
Merged
Conversation
Adds RunContext.session.id, runner-filled from env.sessionId at the startToolRelay call and never service-filled. A missing direct-call binding now names the tool in its error so the model can act on it.
…kit (S2+S3) Endpoint-mode op against POST /api/sessions/streams/header with the session id bound from $ctx.session.id and stripped from the model schema. The name schema requires a non-whitespace character so the tool cannot blank a title. The build-kit overlay emits permission: allow for the rename ops so the first-turn call does not raise an approval card.
…le (T1) Every locally set title is also persisted server-side, so letting the server title win makes a server-side rename (the agent's rename_session tool) visible in the chat without a reload.
… missing target (A2) edit_workflow passed flags= unconditionally, so an edit that carried no flags NULLed the column and dropped the row from the agents list. Flags now pass through only when the edit actually set them. The route returns 404 for a missing or archived workflow and 400 on a path/body id mismatch instead of 200 with count: 0, which the runner would report to the model as a successful rename.
…ispatch (A1+A3)
The direct-call allowlist gains PUT (still four explicit methods, origin
lock and mount check unchanged; PUT serializes the JSON body like POST).
rename_agent is endpoint mode against PUT /api/workflows/{workflow_id}
with the artifact id bound twice from $ctx.workflow.artifact.id (path
and body, so the handler's id-mismatch branch is unreachable) and joins
the default build kit, auto-allowed like rename_session.
…eApp (A4) The API no longer nulls flags an edit does not carry, so the hook stops hard-coding a flag that is wrong for any non-application workflow.
…hod direct-call allowlist (V3) Adds rename_session and rename_agent to the op table, refreshes the stale DEFAULT_BUILD_KIT_OPS enumeration, and syncs the interface inventory: the direct-call method allowlist is GET/POST/PUT/DELETE, an unresolved $ctx token fails the call naming the tool, and the runner augments its dispatch copy of runContext with the live session.id.
…budget (V2) Adds session_header and workflow_header checks that read the stored session stream and workflow artifact rows (pattern-based, never literal, so a correct rename in the model's own words scores as a pass), header seeding for trials that start from a named session or a placeholder agent, and a max_rename_calls budget counted into within_budget — 0 by default so the no-spurious-rename scenario can fail on a call at all.
A second channel scope beside the per-session one: watch:{project}:project
carries low-frequency {entity}-changed frames (ids only, never entity
data) for list pages. The route reuses watch_event_stream and requires
VIEW_SESSIONS and VIEW_WORKFLOWS together — one stream carries both
entity families and the frames carry nothing to filter per permission.
The publisher gains changed(*, project_id, entity, id) with the same
best-effort, one-second-bounded, never-raises contract.
…mmitted writes (W2) set_header publishes one session-changed frame and edit_workflow one workflow-changed frame on the project channel, each after its write commits and each unable to fail it. WorkflowsService gains an optional watch publisher, wired in the router and worker entrypoints.
…d agent lists (W3) Extracts the EventSource lifecycle (visibility open/close, jittered backoff with token refresh, ready revalidation, throttled trailing flush) from useSessionRecordsWatch into useWatchEventSource, and mounts a single project-scoped subscriber: session-changed invalidates the ["session-list", projectId] prefix, workflow-changed invalidates the agents list and the artifact query; ready reruns both to cover frames missed while disconnected. Handlers only invalidate — reconcile stays with the mounted hooks.
…gmentation
A request with no runContext but a live session id now hands the relay
{session: {id}} — the S1 augmentation — so the stale no-run-context
assertion is updated to pin the new shape.
…flow edit endpoints The \S constraint lived only in the LLM-facing tool schemas, so a direct API call could persist a name of spaces — clearing the visible title while the row still held a value. The session header edit now validates on its feature-owned DTO (422), keeping the empty string as the chat rail's explicit clear-title action; the workflow edit handler rejects a present name that is empty or whitespace-only (422), while an omitted name still means no change. The shared Header/ArtifactEdit DTOs are untouched — other git entities reuse them.
…oads complete Uvicorn's graceful shutdown drains in-flight responses BEFORE running lifespan shutdown, and the watch SSE generators never end on their own — so any open browser tab wedged every dev-stack reload permanently (live symptom, seen twice: WatchFiles logs 'Reloading...', the old process stops serving HTTP but keeps draining forever, and the API is dead until a hard docker restart). Since Layout now always holds a ProjectWatch connection, this blocked every dev-stack file edit. The shared watch_event_stream (both the pre-existing per-session route and the new project route ride it, so both had the defect and both are fixed here) now bounds each Redis wait to <=1s and re-checks a module shutdown flag, preserving the heartbeat cadence by counting idle polls; uvicorn's Server.handle_exit — the first thing that runs on SIGINT/SIGTERM, before the drain — is hooked to set the flag (the sse-starlette approach). Client disconnect still releases via generator cancellation, and CancelledError is never swallowed (the finally guards only best-effort teardown). Unit tests pin the shutdown-signal exit, the teardown on that path, and the installed hook actually requesting shutdown.
…ing it to httpx params httpx replaces an URL-embedded query string entirely when params= is passed, so api_call's hardcoded project_id params silently dropped e.g. ?session_id=... and the endpoint 422ed with 'Field required' — observed live as the self-naming benchmark failing at seed/verify (name-02/04 setup, name-01 verification read as no_action) while direct curls worked. api_call now parses the path's query and merges it under the params dict (explicit params kwarg wins, then project_id); a path without a query builds the exact same request as before, and an explicit params= kwarg — previously a TypeError — now merges too. Unit tests pin all three shapes.
…idance, pinned by name-05 Live QA found the split: agents with task-shaped personas call rename_session unprompted (benchmark name-01/03 pass 6/6), but the bare product-default hello-world persona answers and stops in 3/3 organic trials — a single line buried among 16+ tool descriptions loses to 'answer the question'. The standing guidance now rides the platform default persona (_DEFAULT_AGENTS_MD, the surface a fresh UI-created agent actually gets): name the session once the first exchange makes the subject clear, rename only on a genuine topic shift, rename_agent only on an identity/purpose change. The services config.py fallback copy is synced and drift-locked against the SDK builder; benchmark scenario name-05-default-persona-session seeds the persona verbatim (drift-locked by test_default_persona_self_naming.py) as the regression guard for the real new-user experience, with name-04's zero-rename budget still guarding against churn.
dosubot
Bot
added
the
size:XXL
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Repository YAML (base), Organization UI (inherited) Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (6)
🚧 Files skipped from review as they are similar to previous changes (5)
📝 WalkthroughSummary by CodeRabbit
WalkthroughThis PR adds self-naming tools, runner context propagation, workflow and session validation, project-level watch events, frontend cache updates, benchmark scenarios, QA query handling, and related design documentation. ChangesRename tool contracts and catalog
Project watch and frontend updates
Verification and supporting artifacts
Estimated code review effort: 5 (Critical) | ~120 minutes Possibly related PRs
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✨ Finishing Touches 💡 2📝 Generate docstrings 💡
🛠️ Fix failing CI checks 💡
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Member
Author
|
Post-PR real-UI sanity check (live stack, fresh agent created through the composer, working provider key):
This reproduces the benchmark's name-05 one-shot result in the actual product surface, not just at the wire tier. |
Contributor
Railway Preview Environment
Updated at 2026-08-10T13:47:55.519Z |
|
Caution CodeRabbit couldn't update its existing comment. The review summary may be out of date. Error details |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 4
🧹 Nitpick comments (1)
web/oss/src/components/AgentChatSlice/state/sessions.ts (1)
423-425: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick winShorten this comment.
Lines 423-425 add a three-line comment. Replace it with one short line or remove it.
Proposed change
- * - enrich `title`/`createdAt` from the server (a non-empty server title wins: every local - * title is also persisted server-side, and a server-side rename — e.g. the agent's own - * `rename_session` tool — must show without a reload), + * - Use a non-empty server title and server `createdAt`.As per coding guidelines, "Keep in-code comments to at most one short line."
Source: Coding guidelines
ℹ️ Review info
⚙️ Run configuration
Configuration used: Repository YAML (base), Organization UI (inherited)
Review profile: CHILL
Plan: Pro Plus
Run ID: 606015cb-1225-435a-9243-a0593cfc702f
📒 Files selected for processing (56)
.agents/skills/agent-release-gate/resources/qa_matrix_lib.py.agents/skills/agent-release-gate/resources/test_qa_matrix_lib_api_call.pyapi/entrypoints/routers.pyapi/entrypoints/worker_queues.pyapi/oss/src/apis/fastapi/sessions/router.pyapi/oss/src/apis/fastapi/sessions/watch.pyapi/oss/src/apis/fastapi/workflows/router.pyapi/oss/src/core/sessions/streams/dtos.pyapi/oss/src/core/sessions/streams/service.pyapi/oss/src/core/sessions/watch/interfaces.pyapi/oss/src/core/workflows/build_kit.pyapi/oss/src/core/workflows/service.pyapi/oss/src/dbs/redis/sessions/contract.pyapi/oss/src/dbs/redis/sessions/watch.pyapi/oss/tests/pytest/acceptance/sessions/test_stream_header_basics.pyapi/oss/tests/pytest/acceptance/workflows/test_workflows_basics.pyapi/oss/tests/pytest/unit/applications/test_build_kit_overlay.pyapi/oss/tests/pytest/unit/sessions/test_stream_header_merge.pyapi/oss/tests/pytest/unit/sessions/test_watch_endpoint.pyapi/oss/tests/pytest/unit/sessions/test_watch_publish.pyapi/oss/tests/pytest/unit/workflows/test_edit_endpoint_name_validation.pybenchmarks/agent-config-editing/bench_lib.pybenchmarks/agent-config-editing/run_benchmark.pybenchmarks/agent-config-editing/scenarios/09-self-naming.jsondocs/design/agent-self-naming-tools/README.mddocs/design/agent-self-naming-tools/api-design.mddocs/design/agent-self-naming-tools/context.mddocs/design/agent-self-naming-tools/plan.mddocs/design/agent-self-naming-tools/qa.mddocs/design/agent-self-naming-tools/research.mddocs/design/agent-self-naming-tools/status.mddocs/design/agent-workflows/documentation/tools.mddocs/design/agent-workflows/interfaces/cross-service/runner-to-tool-callback.mddocs/design/agent-workflows/interfaces/cross-service/service-to-agent-runner.mdsdks/python/agenta/sdk/agents/dtos.pysdks/python/agenta/sdk/agents/platform/op_catalog.pysdks/python/agenta/sdk/agents/tools/models.pysdks/python/agenta/sdk/utils/types.pysdks/python/oss/tests/pytest/unit/agents/platform/test_op_catalog.pysdks/python/oss/tests/pytest/unit/agents/test_default_persona_self_naming.pyservices/oss/src/agent/config.pyservices/oss/tests/pytest/unit/agent/test_default_agent_template.pyservices/runner/src/engines/sandbox_agent/run-turn.tsservices/runner/src/protocol.tsservices/runner/src/tools/direct.tsservices/runner/src/tools/relay.tsservices/runner/tests/unit/sandbox-agent-orchestration.test.tsservices/runner/tests/unit/tool-direct.test.tsweb/oss/src/components/AgentChatSlice/hooks/useSessionRecordsWatch.tsweb/oss/src/components/AgentChatSlice/state/sessions.pageTitle.test.tsweb/oss/src/components/AgentChatSlice/state/sessions.tsweb/oss/src/components/EntityIdentity/useRenameApp.tsweb/oss/src/components/Layout/Layout.tsxweb/oss/src/components/Layout/ProjectWatch.test.tsxweb/oss/src/components/Layout/ProjectWatch.tsxweb/oss/src/hooks/useProjectWatch.ts
💤 Files with no reviewable changes (1)
- web/oss/src/components/EntityIdentity/useRenameApp.ts
| settled.get("settled", False) | ||
| and len(errors) <= budget["max_tool_errors"] | ||
| and commit_calls <= budget["max_commit_calls"] | ||
| and rename_calls <= budget["max_rename_calls"] |
There was a problem hiding this comment.
🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win
Include rename calls in harness-only attribution.
Line 248 makes rename calls part of the budget. The call at Line 270 does not pass rename_calls to blocked_only_by_harness.
A trial that exceeds max_rename_calls and has only harness errors can be reported as blocked only by the harness. This corrupts the corrected-rate diagnostic.
Proposed fix
- "blocked_only_by_harness": B.blocked_only_by_harness(
- errors, commit_calls, budget
- ),
+ "blocked_only_by_harness": B.blocked_only_by_harness(
+ errors, commit_calls, rename_calls, budget
+ ),Update blocked_only_by_harness to reject records where rename_calls exceeds budget["max_rename_calls"].
Also applies to: 270-272
…t in harness attribution, doc fixes - qa_matrix_lib.api_call keeps the path query as pairs so repeated keys (?workflow_refs=a&workflow_refs=b) survive the merge; new test. - blocked_only_by_harness now takes rename_calls so an over-budget rename trial is never excused as harness-blocked. - api-design.md documents the implemented auto-allow default; qa.md wording.
This was referenced Aug 10, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Context
Sessions in the chat rail keep whatever truncated auto-title the client cut from the first message, and a new agent stays "Untitled agent" until a human renames it. The agent itself is the one participant that actually knows what a session is about, but it had no way to say so.
Changes
Two platform tools that every build-kit agent now carries, auto-allowed so no approval card interrupts the conversation:
rename_sessionsets the current session's name (what the session is about, findable in a list) and description (a one-sentence state recap). The session id is bound server-side from the run context; the model only passes name and description.rename_agentrenames the agent itself overPUT /api/workflows/{id}, with the artifact id bound server-side.What had to change underneath, layer by layer:
rename_sessioncan bind it.PUTjoins the direct-call method allowlist (was GET/POST/DELETE), with the body serialized like POST.edit_workflowno longer nulls flags the caller omitted, and a missing or archived target now returns 404 instead of a success withcount: 0(a deleted agent can no longer be "renamed" successfully).session-changed/workflow-changedbeside the committed writes, streamed to the browser overGET /api/sessions/watch?project_id=(SSE). OneProjectWatchmount in the layout invalidates the session list and workflow queries on each event, and a non-empty server-set name wins over the client's auto-title in the rail.name: ""as the explicit clear-title action; workflows reject empty and whitespace-only alike.Making the model actually use it
Live QA showed the plumbing is not the hard part: task-oriented agent personas call the tools reliably, but the bare default persona a brand-new agent ships with never did (0/3 organic trials). The tool description alone does not drive behavior. The fix is three tight bullets in the default agent persona (the exact surface the failing population gets, drift-locked between the SDK and the services fallback by unit tests): name the session after the first exchange, rename only on a genuine topic shift, rename the agent only when its identity or purpose changes.
Measured by the new
self_namingbenchmark class (5 scenarios, 3 trials each, with enforced rename-call budgets):The two scenarios that define the product experience (05 and 04) pass fully. The residual misses are a general explore-before-acting habit, now permanently measured for future tuning.
Also in this branch
api_call()merges a path-embedded query string into params instead of letting httpx silently drop it (this bug corrupted half of the first benchmark run).The full design workspace (research, API design, plan, QA notes) rides in this branch under
docs/design/agent-self-naming-tools/.Tests
benchmarks/agent-config-editing.What to QA
Full live-QA report (mechanism, scope, model behavior, permissions)
Live QA against
agenta-ee-dev-rel112, driving both the UI and the wire API directly.Tool mechanism, verified solid. Both endpoints persist correctly to Postgres (verified by direct row reads). Both ops are correctly auto-allowed: no approval frame ever appears in the SSE stream, on organic or explicit calls.
GET /api/sessions/watchdelivers realsession-changed/workflow-changedevents on every rename, and the sessions and agents lists live-update without a reload. Whitespace validation verified in all four cases plus a sanity check that normal renames still work.Overlay scope, confirmed. A normal Chat-tab session on an already-created agent carries the full build-kit overlay including both rename ops with
permission: "allow", byte-identical in shape to the onboarding composer's request. Every regular chat session gets the tools.Trigger credentials, not a gap. Trigger-minted tokens carry the creating user's full RBAC role;
EDIT_SESSIONS/EDIT_WORKFLOWSgate both endpoints identically for trigger and interactive auth. The build-kitpermission: "allow"only skips the client-side approval UI, never server-side RBAC.Model behavior. Organic trials with the literal default hello-world persona: 0/3 spontaneous renames before the persona change; the benchmark's task-oriented personas: 6/6 across both tools; explicit-instruction control: 1/1. After the persona change: the table above.
Found and fixed along the way: the server-side whitespace validation gap, a misleading
AmbiguousConnectionErroron an empty vault (#5902), the uvicorn reload wedge, and the harness query-drop bug.