[khushalsonawat]

Review only commit 2d2a794, the one after 1bd1023 (#278's head). This PR is stacked on #278, which is stacked on #271. Their commits show here until they land.

Description

These are three follow-ups left by #245's last review, for the read-only agent reading (TH-8951). None changes anything the platform sees today.

Type of Change

• Bug fix

Changes

• Which jobs the reading refuses. One rule, provider_profile_source in job.py, is now shared by the reading's job check and the build's _load_provider_import_profile. The reading refuses exactly the jobs the build would load a profile for, whether from a provider's API or from the job's own settings.
  • A connect_only job on a connector with no profile (auto, http) is now read, not refused. The platform never sends one.
  • Phone and LiveKit connect_only stay refused, because the build adds their pasted prompt to what it reads.
  • PROVIDER_MANAGED_CONNECTORS moves to job.py and is re-exported from authoring_entrypoint for authoring_runtime_validation.py's lazy import.
  • The refusal message now covers both cases.
  • Tests:
    • an agreement test over 34 connector × mode pairs;
    • an explicit table of the 8 refused pairs;
    • an end-to-end connect_only + auto reading.
• The flaky test. test_nothing_is_saved_after_the_soft_limit_decides waits for the reading's decision instead of a fixed 1.5 s. With the event loop held for 1.4 s, the old test failed 5 of 5 runs. It is now a probe of the reading's cancel of the AI's turn.
• A late save after the decision. _stop_the_rest() is the first line of _read's finally. It cancels every other task on the reading's loop before the reading returns, so a late tool call can't replace contract.json, or write failure.json beside agent.json, after the reading has decided.
  • This isn't reachable with the locked claude-agent-sdk, mcp and google-adk versions, which the hosted image installs. It was real in the reading's own code.
  • Tests:
    • every way out of the reading: finished, soft limit, a session whose close hangs, and a cancel that escapes after the save;
    • a late call woken in the deciding turn;
    • the real submit_contract cancelled at each of its waits, which writes nothing;
    • the reading's Gemini tools running on its own loop.
  • Not covered, by design: a tool that runs on another thread's loop, or one that swallows its cancel. ADK's tool_thread_pool_config applies only to live mode in 2.7.1, which the reading doesn't use.

Known limits

• The reading never sets up the agent's Compose services (older than this PR): it's always told Compose isn't enabled, while the build sets them up, so its view of the agent's dependencies can differ from the build's.

Pending Khushal's OK

• Rewording P10 in the reading's contract: "It also refuses a job the build would load a profile for, from a provider or from the job's own settings: provider_import, or connect_only on vapi, retell, retell_chat, phone or livekit. It refuses these as reading_job_invalid (78), because the build adds that profile to what it reads and a reading never loads it."
• A line to Dileep: "The reading now also reads a connect_only job whose connector has no provider profile (auto, http). The platform never sends one, so nothing changes for TH-8824."

Testing

Each test file ran on its own:

test file	passed
test_agent_reading	190
test_reading_entry	61
test_agent_scan	63
test_call_runner	79
test_harness_architecture	55
test_vertex_gemini_backend	20

The 16 other files that use the moved rule also passed, among them test_authoring_runtime_validation (112), test_agent_type_refusal (651) and test_bundle_author_v2 (323). There were no failures anywhere.

Planted bugs. Each one-token or one-line break was applied in a scratch copy and turned its named test red:

• a connector dropped from the rule;
• the check inverted;
• the strip() removed;
• the soft-limit cancel removed;
• the guard deleted, moved, narrowed or skipped on one path;
• the save swallowing its cancel at either wait;
• a Gemini tool moved off the loop.

The reviewers added 8 plants of their own, all caught except one that is an equivalent change.

Stability: the timing-sensitive tests passed 10 of 10 repeat runs.

Review: a design review, a plan review and one change review round, each by two fresh reviewers. All were CLEAR, and their Mediums are fixed in this commit.

Related Issues

Refs: TH-9454

[cdileep23]

Please check the comments.