feat(opencode): enable exact recoverable conversation rewind

[rynfar]

OpenCode checkpoints could not offer Pylon's rewind actions because its relative numTurns API could not prove retries, compensation or restart recovery. This change uses native immutable forks to enable both restore-files and keep-files rewind through the existing exact rollback saga.

Each eligible completed Pylon turn has a private, immutable snapshot. Admission verifies the current idle native history, account, workspace, incarnation and pending inputs; application selects a fresh verified fork, and compensation selects the untouched source. Native message contents and parent/compaction relationships are checked after ID remapping. Concurrent native completion handlers cannot replace the first captured checkpoint. Original checkpoint targets remain distinct from the current source after compaction and restart. Proofs and snapshot IDs stay private. Historical checkpoints without an exact snapshot remain unavailable. Experimental native plan-file mode and externally managed servers with unobservable runtime flags also remain ineligible for exact rewind; ordinary plan/history/resume continue normally. Native forks cannot copy the external plan file, and Pylon does not alter workspace files or runtime flags to conceal that limit.

The shared service persists selected native cursors before capture/apply/inspect/release receipts, rejects missing cursors and stale same-incarnation writes, and persists compaction state on success/failure before publishing completion. Prime's existing exact saga and provider lifecycle remain intact. Native fork snapshots add storage proportional to the histories retained; no approximate count-based fallback is introduced.

Validation: independent adversarial adapter and service review; 363 focused adapter/helper/service/saga/reconciliation/capability tests; server typecheck and scoped lint. An isolated installed OpenCode 1.18.29 import/fork/read smoke verified full/truncated semantic equality, compaction-tail ID mapping, parentless fork identity, idle/input APIs and unchanged source/workspace. It used synthetic data, no prompts or inherited credentials, and stopped both owned processes. This is native API evidence, not rendered UI or an LLM compaction run. Final-head CI is required before merge.

Part of #526; completes the OpenCode portion of the #513 exception for upstream fd5553f1afcef4f410a067687faa991743b5034c and efccda9ac9230db22b36990cffabdad218fa41b0. Codex and Claude follow separately. The compact ledger records the already merged sidebar, public preview, mobile Antigravity and Claude quota dispositions; the cycle cursor remains unchanged pending final accounting.

Preview workflow closure additionally depends on the independently reviewed checkout repair in #532; the owning issue tracks its actual post-merge execution.

Model: GPT-6. Harness: Codex.

[vercel]

Deployment failed for project pylon-marketing with the following error:

Resource is limited - try again in 24 hours (more than 100, code: "api-deployments-free-per-day").

Learn More: https://vercel.com/pylon-code?upgradeToPro=build-rate-limit

[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	14.1 KiB	14.1 KiB	+4 B (+0.0%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.2 KiB	7.2 KiB	−5 B (−0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.9 KiB	6.9 KiB	+9 B (+0.1%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	58.8 KiB	58.8 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	10	10	0 (0.0%)	21	✅
Claude	Total thread wire	14.1 KiB	14.1 KiB	−12 B (−0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.2 KiB	7.2 KiB	−6 B (−0.1%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.9 KiB	6.9 KiB	−6 B (−0.1%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	59.7 KiB	59.7 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: d4e599b · PR result: 42e4959 · 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.

[vercel]

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

Project	Deployment	Actions	Updated
pylon-marketing	Ready	Preview	Sep 12, 2026 5:13pm UTC