Repository navigation
✨ meetings: copilot and omp completion profiles, gateways/vllm/llm-d, operator-token Claude (#18549) - #18556
Conversation
… operator-token Claude, refusal reasons (#18549) Agent-backend mode now covers every backend the agents' picker offers: - copilot (1.0.88) and omp (18.20.0) answer through one tool-less CLI per call with the agents' sign-in, prompt on stdin, usage charged to the meeting ledger; copilot runs in a throwaway home. - vllm, llm-d and named model gateways map onto the OpenAI-compatible path (gateway key, default model and CA bundle). - A hive with only HIVE_CLAUDE_OAUTH_TOKEN passes it to the meetings CLI as CLAUDE_CODE_OAUTH_TOKEN, as the agent launcher does. - codex, gemini, watsonx and the other CLIs are refused at validation, in PUT /api/meetings/model and in the Settings picker with the reason. Signed-off-by: Andy Anderson <andy@clubanderson.com>
Open PR overlap checkBase branch: Conflicts are checked with every PR merged onto the current tip of
|
Mutation testinggremlins found 3 surviving mutant(s) in changed files: the code ran under tests, but changing it did not fail any test.
7 mutant(s) in changed files are not reached by any test (NOT COVERED). A maintainer may apply the
ℹ️ 4 mutant(s) on added lines were SKIPPED because gremlins v0.6.0 only honours the first
Verdict: |
There was a problem hiding this comment.
Decision needed: @clubanderson should confirm that the copilot --available-tools hive-meetings-no-tools launch really runs with no tools on the pinned 1.0.88 CLI before this hands it a live Copilot token.
Read the diff and the launch/parse code in src/pkg/meetings/meetllm/cli.go and src/cmd/hive/meetingswire.go. I found no code defect, and the per-call temp dir, stdin prompt delivery, and error and empty-reply handling look correct to me.
One open question, in the credentials risk class. This change gives the meetings CLIs the agents' sign-in (COPILOT_GITHUB_TOKEN, the operator Claude token, the omp broker variables) while they answer from attendee-controlled meeting text. For copilot, "no tools" is expressed as an allow-list of a name that does not exist (copilotNoToolSentinel, cli.go). The PR body pins 1.0.88 but reports its measurement on 1.0.109. I could not verify from the tree that the sentinel disables every tool on the pinned version. If it does not, an injected prompt could reach a CLI that holds the token. The refusal of codex for exactly this reason shows the concern is already recognised here. A short check on 1.0.88 (a tool-use prompt with the sentinel) would settle it.
I have not approved or blocked anything.
— hive: agent=reviewer backend=claude model=claude-sonnet-5-5 claude=2.1.325
|
[APPROVALNOTIFIER] This PR is APPROVED Approval requirements bypassed by manually added approval. This pull-request has been approved by: The full list of commands accepted by this bot can be found here. The pull request process is described here DetailsNeeds approval from an approver in each of these files:Approvers can indicate their approval by writing |
Follows #18528. Every backend the agents' backend picker offers is now either selectable for meetings with a tested completion profile, or refused with a specific, documented reason.
Selectable
--available-tools <none> --disable-builtin-mcps --no-custom-instructions --no-ask-user --no-auto-update --output-format json --usage-output-file <tmp>. The prompt goes in on stdin, not argv.HOMEandCOPILOT_HOMEpoint at the call's temp dir, so nothing from the agents' home loads and the session log is deleted with the call. Sign-in isCOPILOT_GITHUB_TOKEN, or the dashboard Copilot token read through the keystore, the same way agents get it. The reply comes from the lastassistant.message, and usage is input plus output tokens from the usage report.-p --mode json --no-tools --no-lsp --no-extensions --no-skills --no-rules --no-session --no-title --no-context-files --profile hive-meetings --system-prompt <fixed>. The prompt goes in on stdin. Sign-in comes from the auth broker variables. A model is required. The reply is the last assistantmessage_end, and usage is itsusage.totalTokens.HIVE_VLLM_ENDPOINT,HIVE_LLMD_ENDPOINT), with no key.default_modelandca_bundle. A newHIVE_MEETING_LLM_CA_BUNDLEcarries the CA bundle tohive meeting-llm, and verification is never turned off. Gateways show in the picker under their own names.HIVE_CLAUDE_OAUTH_TOKENor its file) reaches the meetings CLI asCLAUDE_CODE_OAUTH_TOKENin its environment, as the agent launcher gives it to agents. Hive never logs it or writes it to disk.Refused (validation,
PUT /api/meetings/model, and the picker with the reason)codex (no flag removes its shell tool; read-only commands could read the sign-in), gemini (not in the image, no tool-off flag, consumer accounts ended), agy and bob (no tool-off flag), goose, aider, cline, pi, opencode, kilo, muse, openhands and omnigent (no verified tool-less launch), watsonx (needs an IAM bearer and a project header), and any name that is not a configured gateway. A test checks that every name in
CLIBackendsandInferenceBackendsis either supported or refused.The claude-cli path and Betty's behavior are unchanged, apart from the shared code. Spek on the warm process (optional in the issue) is not included.
Latency is listed per backend in
src/docs/meetings.md. On copilot, measured on 1.0.109 (the newest available locally), a call took about 4.5 s from launch to the model's response. omp latency has not been measured against a signed-in CLI.Fixes #18549
🤖 Generated with Claude Code