docs: record platform runtime adoption checkpoint

[thymikee]

Summary

Record ADR 0019 checkpoint outcome as revise.

• Preserve the accepted, coherent devices, logs, and network cutovers.
• Record passing architecture, behavior, lifetime, laziness, startup, and coupling evidence.
• Record the hard package-size failure and prohibit broader migration pending an explicit packaging/budget decision and clean checkpoint rerun.

Stacked on #1702. Tracks #1696. Follow-up decision: #1704.

Validation

• pnpm check:affected --run (3,968 tests; 87.70% changed-line coverage)
• controlled same-host pnpm build + node scripts/size-report.mjs --startup-runs 15 baseline/checkpoint comparison
• final adversarial claude -p review; confirmed network line-offset P1 fixed in refactor: extract platform network runtime #1702, no remaining P0/P1 findings

[thymikee]

Review: changes requested

Reviewed at exact head ecb75aa0723f0a1bb212c46145207344bb1ab249.

P1 — the checkpoint decision records disproven behavior evidence. ADR 0019 says checkpoint 99f5af1 passed behavior/architecture evidence, treats the logs/network cutovers as accepted coherent units, and says final review left no P0/P1 findings. That is false at the recorded checkpoint: #1701 a108975 still has the stale owned-PID-marker path where terminal recovery leaves the marker and the next logs start wedges; #1702 99f5af1 still bypasses the request-scoped Apple tool provider during iOS simulator recovery and regresses Android optional-logcat recovery failure handling. Neither has a fixing commit, and exact-head changed-route live evidence remains absent.

Please revise the ADR to record these correctness/evidence blockers alongside the package-size failure, and require them to be closed before the checkpoint can rerun or record continue. Do not call the cutovers behavior-passing/coherent until then.

The revise status, size decision in #1704, and broader-migration prohibition are otherwise correct.

All three docs-only checks are green and the PR is mergeable/CLEAN; no label applied.

[thymikee]

Re-review: not ready. Head remains ecb75aa, its stacked base has moved to corrected #1702 head d73bdb4, and GitHub reports DIRTY/conflicting. The former #1701/#1702 correctness and live-evidence blockers are now resolved in their new heads, but this ADR still presents old checkpoint 99f5af1—which contained those defects—as behavior-passing/no-P0/P1. Rebase and resolve the stack, then either record 99f5af1 as a failed/superseded checkpoint or update the exact corrected checkpoint SHA and evidence after a clean full rerun. Retain revise while #1704’s package-budget decision remains open. No ready label applied.

[thymikee]

Exact-head re-review at ecb75aa0: not ready. The branch is still DIRTY/CONFLICTING on the obsolete stack: its merge-base with current #1702 8fd0d111 is the original baseline 44c298d7f, so the checkpoint commit has not been rebased onto the corrected network unit. Rebase the docs-only change onto current #1702, resolve the ADR conflict, and obtain fresh authoritative CI.

The ADR checkpoint is materially stale, not just SHA-stale: it still treats old 99f5af1 as behavior/ownership clean with no remaining P0/P1 and records superseded package, layering-test, affected-test, and evidence counts. The corrected #1701/#1702 stack includes lifecycle/admission, capture-kit, and recovery changes, so record the corrected exact checkpoint SHA and fresh gates/evidence; if retaining the old run, mark it superseded rather than current acceptance evidence.

#1704 correctly remains revise, and the cumulative locked +3% condition is still unresolved. The latest cumulative same-baseline evidence already exceeds all four limits, while current #1702 adds size. Provide a clean same-host 15-run 44c298d7f → corrected-stack measurement and document the accepted revised bound/checkpoint; per-PR deltas cannot replace the cumulative checkpoint. No ready label applied.

[thymikee]

The checkpoint PR is rebased and rewritten at ffc077e25cd2ab7cd9ddfe3b1100efb59ba18619, on final network head 457fafe6399a95a4ddbfac57f02b3a7fe4157a54.

The ADR now records:

• original cumulative denominator 44c298d7f3a0ef84bc47f34c54d88b6c9eeb0df2 (not the moving post-refactor: extract platform device inventory runtime #1699 base);
• merged devices c06bed9f, logs 188795386, and network/checkpoint 457fafe63;
• the old 99f5af1/d73bdb4 checkpoint as superseded and not behavior-passing;
• the revised budget rationale from ADR 0019: resolve the platform-runtime package-size checkpoint #1704: reliability, cloud streaming, and substrate work—not a new observability layer;
• @agent-device/capture-kit ownership and the contracts implementation ratchet;
• the ADR §5 admission-ledger amendment;
• recordings onto DurableCaptureResource as the next post-checkpoint command unit.

Cumulative same-baseline size evidence is explicit: raw +4.696%, gzip +5.380%, tar +3.731%, unpacked +3.499%. Startup/help remain improved. The exact-head Android logs rollover evidence is also recorded (PID 10952 → 11455, marker rotation, same app.log, successful stop/recovery).

Validation on this exact rebased head: pnpm check:affected --run passed, including 799 files / 6,424 tests, 87.34% changed-line coverage, 131 layering/model tests plus the production scan, fallow, build/package checks, provider integration, integration-progress, and replay compatibility.

[thymikee]

Exact-head re-review at ffc077e2: not ready. The ADR now correctly supersedes the defective 99f5af1/d73bdb4 checkpoints, records the corrected stack SHAs, preserves the original 44c298d7f cumulative denominator, and accurately carries #1704’s revised measured bound and rationale. The size figures remain applicable to final 457fafe6, whose post-measurement delta is package-neutral layering policy/test code, and the next authorization is properly limited to recordings on DurableCaptureResource.

The continue decision is still premature because the hard behavior checkpoint lacks final-rebased practical evidence. #1701 explicitly leaves exact-head iOS logs doctor → start → marker → relaunch → stop → close unverified, and #1702 still has no exact-head iOS and Android logs clear --restart → generate traffic → network dump → stop/close recovery proof after the lifecycle and Apple-host rebase. Keep the checkpoint at revise until those production routes are observed (recording Harmony unavailability as residual risk if necessary), or attach the evidence before authorizing recordings. The stacked #1702 iOS smoke also needs its currently pending green rerun after a known-shape main-thread timeout. No ready label applied.

[thymikee]

Readiness follow-up at ffc077e: the practical-evidence blockers from the prior comment are now closed.

• Final logs head 1887953 has exact-head iOS simulator proof for logs doctor, start, marker, app relaunch, second marker, stop, path, and close: refactor: extract platform log runtime #1701 (comment)
• Final network head 457fafe has exact-head iOS and Android proof for clear --restart, controlled WebView traffic, active network dump, stop/close, plus honest inactive recovery outcomes: refactor: extract platform network runtime #1702 (comment)
• The only remaining stacked CI red was an unrelated XCTest main-thread timeout in the fixture deep-link confirmation. Targeted attempt 2 passed every iOS smoke phase: https://github.com/callstack/agent-device/actions/runs/31401871977/job/93508441731

The ADR already preserves the original 44c298d denominator, records the revised cumulative bound and rationale from #1704, supersedes the defective checkpoints, and limits the next unit to recordings on DurableCaptureResource. #1703 is CLEAN/mergeable, its checks are green, and there are no unresolved review threads.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-08-10 15:59 UTC