docs(upstream-review): track T3 watch items where the review workflow reads them

[rynfar]

Problem

Issue #114 carries an upstream-watch ledger that says to recheck its entries "during Prime Agent upgrades and T3 upstream reviews". The review-t3-upstream skill reads .agents/upstream-review.md, which had zero references to #114 in either direction. The duty was written in an issue body that no workflow opens, so it was never enforceable.

There is a structural reason this could not have been caught by the existing process either: three of #114's watched items are unmerged upstream work. A review bounded by <cursor>..t3code-upstream/main can only see commits, and a pull request that never merges produces no commit. Phases 2 and 3 of the skill are blind to that class of work by construction.

The gap has already bitten. Checking the five watched T3 items against upstream today:

Item	State
#5760 OpenCode todowrite → task sidebar	open
#8097 restore composer tasks after relaunch	open
#6573 authenticated cross-thread MCP tools	closed unmerged 2026-08-28
#7966 server-side scheduled tasks (issue)	closed 2026-08-28
#7986 server-side scheduled tasks (PR)	closed unmerged 2026-08-28

Three changed state on 2026-08-28 — the same day as the last T3 upstream review — and #114 still described them as live upstream signals. Nothing was adopted wrongly, because none of the watched work ever merged, but nothing noticed either.

Fix

Give each fact one owner, and put the T3 facts where the workflow provably looks.

• .agents/upstream-review.md gains an ## Upstream watch list: upstream work that never enters a review range, keyed WATCH-n, each with an Owner issue, a current signal, a checkable Revisit when, and the response Pylon intends to evaluate. WATCH-2 and WATCH-3 are rewritten around the closures — the implementations Pylon declined to cherry-pick no longer exist, so both now watch for a replacement with a 2026-11-01 floor rather than describing a dead PR.
• This is deliberately not folded into the ## Deferred register. Register entries were candidates in a batch, carry an upstream SHA, and end in adopt or skip. Watch entries have no SHA and may never land; conflating them would break both lifecycles.
• review-t3-upstream/SKILL.md now reads and maintains the list: preflight checks it exists (phase 1), phase 2.5 re-evaluates it with gh pr view --json state,mergedAt and treats a closed-unmerged PR as a real signal, phase 4 reports state changes, phase 5 maintains it, and the handoff is incomplete without it.
• Prime provider parity: tasks, automation, and upstream watch #114 keeps its three Prime Agent rows — a T3 review cannot evaluate PrimeIntellect-ai state — and points at the WATCH- ids instead of duplicating them. Updated in the same change; no table is mirrored in two places.

Verification

vp fmt --check passes on both files. No code paths touched.

Not included

docs/internals/prime-agent-daemon-parity.md is called authoritative by #114, which asks that both be updated when a linked PR lands. It has no mention of the user/delegate/external wait distinction that PR #137 shipped. That is a content gap in a different ledger and belongs in its own PR.

---

Claude Opus 5 in Pylon.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ The exact PR base did not have a successful artifact. Baseline uses the latest successful main measurement shown below.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.3 KiB	−18 B (−0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	−1 B (−0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.4 KiB	−17 B (−0.3%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.5 KiB	55.5 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	9	9	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.3 KiB	+16 B (+0.1%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+1 B (+0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.4 KiB	6.5 KiB	+15 B (+0.2%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	56.4 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	10	10	0 (0.0%)	21	✅

Baseline: 750d503 · PR result: f0e015d · 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: 109.4 KiB
• Claude decoded thread snapshot: 110.1 KiB

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