[amilandi]

Problem

Capability-first ADK entry points allocate and persist a session ID without emitting adk.session.start. A later explicit start_session() previously treated the allocated ID as evidence that a start had already been emitted and skipped it. Capability usage could therefore appear in External DA charts while sessions and active installs remained zero.

Changes

• Separate session-ID allocation from persisted start-event recording. Ensure starts in shared single-event and validated client-batch dispatch, rather than relying on individual skill wrappers.
• Match starts to the activity's install, tenant, agent type, and telemetry project. An earlier anonymous, CEA, other-install, or dev-project start cannot suppress the start required by another dashboard cohort.
• Mark a context recorded only after delivery or successful buffering; retry starts that could neither be sent nor buffered. Preserve opt-out, async dispatch, the 30-minute inactivity window, and existing event dimensions.
• Reserve session starts under short local-state locks, buffer follower activity while a start is pending, and expire abandoned reservations. Atomically detach retry-buffer snapshots before collector I/O so slow flushes do not block or lose concurrent appends. Live events join the queue while an older snapshot is in flight, preserving start-before-activity ordering.
• Document the reporting semantics and best-effort delivery limits. Active-user charts count distinct installs, not identified people. A session spanning attribution contexts may have a start in each; deduplicate session IDs when combining cohorts.

Verification

381 targeted tests passed across the SDK, authentication, Workday telemetry, FlightCheck telemetry, and telemetry guard suites. Regression coverage includes capability-first starts, matching External DA attribution, legacy/allocated session files, timeout rollover, consent changes, buffered replay, failed buffering, async responsiveness, cohort changes, and simultaneous fresh capability-shim subprocesses. Ruff and diff whitespace checks pass.

Merge order

PR #458 is merged, and this branch is rebased onto it with both tenant-classification and session-reporting documentation/tests preserved. Rebase #436 after #449 so its LPC emitter changes are applied to this finalized transport.

Scope

Prospective emitter fix only: existing clones must pull the change; historical missing session-start events are not backfilled. No Aria cube or FlightCheck outcome-event schema changes. Dashboard test-tenant exclusions were saved separately and are not repository changes.

[nkemms]

• This change separates session allocation from start recording and adds useful regression coverage, but the new cross-process locks are held across collector network calls.
• Do not merge: 0 CRITICAL and 3 MAJOR findings. Two reproduced lock-timeout paths violate ordering or lose telemetry; active PRs 458 and 436 also have textual merge conflicts with this branch.
• Targeted validation passed: 301 tests, Ruff, and diff whitespace checks.

Verdict: Do not merge until network I/O is moved outside the session-start and buffer file locks, contention tests cover waits beyond the lock deadline, and merge order is coordinated with PRs 458 and 436.

[MAJOR] F-3 — Active telemetry PRs have textual merge conflicts
Owner: Dev | Confidence: High | Impact: Medium

git merge-tree reports conflicts with PR 458 in TELEMETRY.md and tests/test_adk_telemetry.py, and with PR 436 in adk_telemetry.py. Both siblings are open and non-draft. PR 458 is current and shares the same author; PR 436 adds LPC operation telemetry to the same emitter.

Fix: Choose a merge order now. Rebase whichever PR merges second and preserve both telemetry changes and their tests.

[amilandi]

Addressed the concurrency findings in 2412c309 and selected the merge order for F-3: #458 → #449 → #436. PR #449 will be rebased after #458 merges, preserving both TELEMETRY.md/test changes; #436 should then rebase onto the finalized transport.

The new contention tests hold collector calls for 6.2 seconds, exceeding the former five-second lock deadline. Validation: 368 passed; Ruff and whitespace checks clean.

[amilandi]

Resolved the merge conflicts by rebasing onto current main after #458 merged. Preserved both the internal-test-tenant attribution section/tests and the session-start transport section/tests. Rebased head: 2ef0a9c7. Validation: 381 focused tests passed; Ruff and diff whitespace checks are clean.

[Dawn Jeong (daeunJe0ng)]

[Copilot-review] Request changes at 2ef0a9c762de7ad6d7eddb70df15d622d75af364. Separating session allocation from recorded starts and moving network I/O outside state locks are sound improvements. The snapshot-ownership race can still lose a buffered start. The existing normal-exit loss thread remains applicable: #449 (comment). Cross-context replay also violates the documented ordering guarantee. Static review only; no tests, builds or live telemetry were run.

Addressed in 0067784

[Dawn Jeong (daeunJe0ng)]

[copilot-review] Approving d3035ad9 with the three unresolved findings below treated as follow-up work. The three earlier blockers are fixed in their demonstrated forms. The session-expiry recovery issue remains unresolved; this approval does not mean it was fixed. Static review only; no tests, builds or live telemetry were run.

[Dawn Jeong (daeunJe0ng)]

[copilot-review] Request changes at 31ef143e8fbdd6516a104e98fd926752a7118662. The shared session-start enforcement is the right design, and the previous rollover-evidence, legacy-buffer-format and buffer-budget findings are fixed in their demonstrated forms. One blocking retry-selection defect remains: eligible buffered activity can be starved indefinitely while later events keep uploading. Completion-ledger growth is a separate nonblocking follow-up. Static review only; no tests, builds or live telemetry were run.

[Dawn Jeong (daeunJe0ng)]

[copilot-review] Approve at c5b58f09f87384d46b80e9de8f2fa9f1172421d1.
The previous retry-starvation blocker is fixed: successful raw-snapshot finalization now promotes eligible waiting followers.
Completion history is bounded, and recovery after evidence eviction preserves the original follower attribution and start-before-activity ordering.
I withdraw the previous findings for their demonstrated triggers, including the completion-history follow-up.
The inspected regression tests cover the public-emission retry path and eviction recovery; no remaining Blocking or Should-fix defect was established.
Static review only; tests, builds, and live telemetry were not independently exercised.