fix(prime): reconcile results for observed native tool calls

[rynfar]

Native Prime can publish a completed tool result in a transcript snapshot before its message-completion event. Pylon rejected that delta even when it had already observed the exact tool call, failing the first real turn after MCP tool discovery.

The adapter now reconciles unique results whose IDs and names match the current correlated prompt’s attributed live tool calls. Exact transcript overlap, complete replay, current connection/proof epoch, delivered lifecycle continuity, and no queued input remain required. Unrelated calls, changed names, duplicate or previously recorded results, and mixed assistant deltas remain rejected.

The runtime also suppresses an exact late tool-result completion already published in the current attachment’s bounded correlated snapshot. This prevents duplicate transcript progress without hiding changed results or affecting provisional worker snapshots.

Validation: both the missing-result regression and delayed-completion regression fail before their fixes; 432 adapter/runtime tests pass. Coverage includes attribution through an intermediate snapshot, duplicate/unknown results, proof/lifecycle failures, and changed result content/timestamps. Server typecheck and scoped lint pass. The actual managed stable build completed the previously failing flow in 15.5 seconds with a ready checkpoint and no alerts; the final corrected real tool flow also passes. The actual restart proof passed 2/2 (116.74s) with failure-only diagnostics; an earlier run had an unexplained second-checkpoint error, retained in the evidence. No test assertion was weakened. Final checkpoint review passes at 3eb0769; full web/mobile acceptance continues in #114.

Model: GPT-6, Codex harness.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
pylon-marketing	Ready	Preview	Sep 12, 2026 3:18am UTC

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	14.0 KiB	13.9 KiB	−145 B (−1.0%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.2 KiB	7.2 KiB	+1 B (+0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.8 KiB	6.7 KiB	−146 B (−2.1%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.8 KiB	58.0 KiB	−866 B (−1.4%)	66.4 KiB	✅
Codex	Live turn messages	9	8	−1 (−11.1%)	21	✅
Claude	Total thread wire	14.0 KiB	14.0 KiB	−28 B (−0.2%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.2 KiB	7.2 KiB	+11 B (+0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.8 KiB	6.8 KiB	−39 B (−0.6%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	59.7 KiB	59.6 KiB	−44 B (−0.1%)	66.4 KiB	✅
Claude	Live turn messages	9	8	−1 (−11.1%)	21	✅

Baseline: c6a8b64 · PR result: 3eb0769 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 115.6 KiB
• Claude decoded thread snapshot: 116.3 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[rynfar]

Actual managed stable native conversation after the reviewed correction: tool use and a final answer complete with a ready checkpoint and no error alert. The earlier failed first-turn recording and diagnostic sequence are retained with #114 acceptance evidence.