[congvc-bot-x99]

Status

Agent-side verification is complete. Ready for Cheese's merge review. This is not merge or production-release authorization.

Conditions

• No production rollout. A new response field needs the normal reviewed release before live use.
• Current head a1000970d90bb6dcf87b8168a184e5a863b7a7b5, base main at 93ac2140b9ce58564b71b8b48793dcda811d60da. The head is main merged into the branch; the diff against main is still four files.
• Independent different-lab review: c00-hermes-devops (OpenAI gpt-6.1-sol), independent of the Sonnet/Anthropic author, signed PASS on the current head/base. The signed delta report includes fresh OCR delegation and manual supplements. It covers the unchanged feature patch and main integration, not every imported main commit independently. The earlier full-feature review covers head 59bf08b6cf2bb4047263c0f04dc01bdb7b15bd99.
• No direct production database query, no C00 write, no server restart, no CLI replacement.

Change

multica agent tasks and multica issue runs (full log, JSON) now return claimed_model on every run. The server reads it from the stored agent_task_queue.claimed_model column. It does not read the agent's current model.

• null: nothing stored (unclaimed, or the claim could not record it).
• "": an empty string was stored.
• Key absent: the server predates the field.
• Claim time is the existing dispatched_at. Task ID, agent ID, issue ID, workspace ID and status were already returned.
• issue runs --siblings is a compact coordination read and is unchanged.

No new endpoint and no CLI change: the CLI prints the server JSON as is. Existing workspace-membership and private-agent checks stay on both reads. No prompt body, provider environment or token is added.

Remaining gap: claimed_model attests what the daemon was told to run at claim. It does not show which upstream model or provider served the run. That stays NOT-VERIFIED. No alias, lab inference or historical backfill is added.

Files: server/internal/handler/agent.go, server/internal/handler/daemon.go, server/internal/handler/task_claimed_model_response_test.go, apps/docs/content/docs/cli.mdx.

Evidence

Red, before the fix (go test ./internal/handler -run TestTaskReads, throwaway pgvector Postgres 17 on X99):

    task_claimed_model_response_test.go:91: task 6ad283cd-4e85-46e3-b347-34aa08e1569a: claimed_model key missing
    task_claimed_model_response_test.go:91: task 87d10190-77b6-41a3-9e53-153a9896032b: claimed_model key missing
    task_claimed_model_response_test.go:91: task ed3085d1-1849-4ed5-a82e-26fe8a0767ab: claimed_model key missing
    task_claimed_model_response_test.go:91: task ed3085d1-1849-4ed5-a82e-26fe8a0767ab: claimed_model key missing
    task_claimed_model_response_test.go:91: task 6ad283cd-4e85-46e3-b347-34aa08e1569a: claimed_model key missing
    task_claimed_model_response_test.go:91: task 87d10190-77b6-41a3-9e53-153a9896032b: claimed_model key missing
--- FAIL: TestTaskReadsExposeClaimedModel (0.05s)
    --- FAIL: TestTaskReadsExposeClaimedModel/agent_tasks (0.01s)
    --- FAIL: TestTaskReadsExposeClaimedModel/issue_runs (0.00s)
--- PASS: TestTaskReadsClaimedModelKeepAuthorizationBoundaries (0.05s)
    --- PASS: TestTaskReadsClaimedModelKeepAuthorizationBoundaries/agent_tasks/non-member (0.00s)
    --- PASS: TestTaskReadsClaimedModelKeepAuthorizationBoundaries/agent_tasks/plain_member (0.00s)
FAIL
FAIL	github.com/multica-ai/multica/server/internal/handler	0.166s

The authorization test passed before the fix too: it guards the existing boundary and does not prove the new field.

Green, after the fix: TestTaskReadsExposeClaimedModel (both reads, string / null / empty) and TestTaskReadsClaimedModelKeepAuthorizationBoundaries pass. A private-agent plain member and non-member get 403/404 with no claimed model in the body. A non-member cannot read a foreign-workspace issue's runs. The full ./internal/handler package passed (244s) on the pre-merge head. After merging main, go build ./... and the targeted tests passed.

Not run locally: server/cmd/multica tests. The author reported workspace tests there fail because of the daemon-task marker in the managed workdir. This PR does not touch the CLI. Current-head CI runs that package and passes (19.817s).

CLI examples (also in apps/docs/content/docs/cli.mdx):

multica issue runs CHE-1962 --output json | jq '.[] | {id, agent_id, issue_id, workspace_id, status, dispatched_at, claimed_model}'
multica agent tasks <agent-id> --output json | jq '.[] | {id, claimed_model}'

Correction after review (P2)

The daemon claim response is built before the claim writes claimed_model, so it returned a stale null. It now reports the value this claim stores (daemon.go), including "" when the agent model is empty. Foreign-workspace test now seeds a secret task and asserts the body does not leak it.

Red, before the correction (go test ./internal/handler -run TestClaimResponseReportsStoredClaimedModel):

    task_claimed_model_response_test.go:174: claim response claimed_model = <nil>, want the stored "claim-response-model"
    task_claimed_model_response_test.go:174: claim response claimed_model = <nil>, want the stored ""
--- FAIL: TestClaimResponseReportsStoredClaimedModel (0.19s)
    --- FAIL: TestClaimResponseReportsStoredClaimedModel/model_set (0.11s)
    --- FAIL: TestClaimResponseReportsStoredClaimedModel/agent_model_empty (0.07s)
FAIL
FAIL	github.com/multica-ai/multica/server/internal/handler	0.268s
FAIL

Green: that test, TestTaskReads* and TestVoteGate_ClaimRecordsModelOnTheTask pass. The full ./internal/handler package passed (233s) on the corrected head.

CI on the current head

Verified pre-readiness rollup on a1000970d: 16 SUCCESS, 6 SKIPPED, 0 failed, 0 pending. This is 14 successes by unique check name, with duplicated successful Changelog-line and close-target events. SKIPPED is not PASS. No waiver.

CI run 39922117900 passes. backend-tests and backend pass. The handler package passes (144.847s), including task-read, claim-response and authorization regressions. The receipt package passes (1.425s). The earlier red TestObserve_EditAfterSettledCreateIsAdmittedAgainstRealBudgetService was the CHE-1578 budget-clamp flake, fixed on main by PR #295.

Working PR for CHE-1962 (the child of CHE-1950 that delivers this read): #357

Writing receipt: passive-voice (five) and semicolon (one) advisories are true positives. The unchanged-state warning and the synonym-rotation warning are false positives.

Changelog: none

Closes CHE-1962