TL;DR

Retrospective systematic audit across all 9 sync batches (v2026.2.25 through v2026.3.22 — fork's full upstream sync history) identified 32 upstream fix commits whose new files exist at upstream/v2026.3.22 but never landed in fork main — the same pattern as #2362 (@tloncorp/api → tlon-api.ts), fixed in #2402.

Filing as a single umbrella to (a) document the full scope discovered, (b) enable batch-porting when convenient, (c) track the known gaps explicitly so they don't re-accumulate. The INCLUDE Reconciliation Gate committed to HQ prevents this pattern for future batches.

Background

Sync to v2026.3.22 (PR #2400) missed the upstream fix ac850e815b that replaced @tloncorp/api git-dependency — the fix was in-range but its new file (extensions/tlon/src/tlon-api.ts) was silently reverted during D.3 build stabilization. Root cause analysis (memory record) revealed the "asymmetric revert-vs-extract split" anti-pattern: stabilization logic recorded only new-upstream-file deletions, not pre-existing-file reverts. Same pattern may exist in earlier batches.

Methodology

1. Enumerate sync batches in fork main (git log --grep="^sync:") — 9 batches from v2026.2.25 to v2026.3.22.
2. Extract upstream range from each sync commit's Upstream-range: trailer.
3. Filter to fix: commits in each range (high-signal for user-impacting work).
4. Find A-status files introduced by each fix commit — 551 across all batches.
5. Filter candidates: file not in fork main, not classified as EXCLUDE*/EXCLUDE-GUT/EXCLUDE-STUB, not under an EXCLUDE directory rule — 55 candidates.
6. Systematic validation: for each, check presence at upstream/v2026.3.22.
   • YES → real gap (fork missing what upstream currently has)
   • NO → correctly absent (upstream self-deleted within sync range)
7. Semantic refinement: exclude paths under fork-gutted parent dirs that have no explicit rule (src/agents/pi-extensions/, extensions/qwen-portal-auth/ — dirs absent from fork entirely).

Result: 32 real gaps, 23 correctly-absent (20 upstream-self-deleted + 2 semantically gutted + 1 rebranded).

Validation artifacts: .tmp/systematic-validate-*/ (TSV + markdown). Audit scripts in HQ scripts/audit-lost-files.sh and scripts/audit-fix-commit-gaps.sh.

Findings — Full list (32 gaps → 27 real, 5 false positives moved below)

Updated 2026-04-19 during the hq B1–B10 cleanup pass. The audit methodology was hardened to filter files deleted by fork gut(*) commits (hq@fbdc14b) — 5 entries below are false positives architecturally gutted from the fork pre-B10. They've been moved to the § Rejected as false positives section and given EXCLUDE-GUT disposition rules (hq@b2e4633). Additional P3 test entries may also be fork-gutted area-related — still under triage (Step 3d).

🔴 Production files (4)

🟡 Test files (24)

Rejected as false positives (27 — documented for audit trail)

• 20 upstream-self-deleted — upstream added-then-refactored within sync range (e.g., src/gateway/system-run-approval-env-binding.ts added 2026-03-12 19:59, refactored 12 min later 20:11; src/infra/outbound/built-in-channel-adapters.ts added then refactored 31 min later)
• 6 fork-gutted by 1aae7daa00 (Playwright gut gut: remove Playwright automation framework from browser subsystem (MBP audit) #2304/gut(browser): remove Playwright automation framework #2330) — files gutted by the fork pre-B10; audit's semantic refinement did not cross-reference fork gut() commits. Now EXCLUDE-GUT in disposition.tsv (hq@b2e4633):
  • src/browser/output-atomic.ts — was fix(browser): harden writable output paths — B10 audit #2416 (P1), closed 2026-04-19
  • src/browser/session-tab-registry.ts — was fix(browser): close tracked tabs on session cleanup — B10 audit #2417 (P1), closed 2026-04-19
  • src/browser/session-tab-registry.test.ts — was in P3 test batch
  • src/cli/browser-cli-actions-input/shared.test.ts — was in P3 test batch
  • (also added but not in original Audit: upstream fixes silently lost in sync (v2026.2.25 → v2026.3.22, 32 gaps) #2406 list: src/agents/tools/browser-tool.actions.ts, src/browser/{form-fields,safe-filename,routes/agent.act.download,routes/agent.act.hooks,routes/output-paths}.ts)
• 1 fork-gutted by e2cf15c94d (stubs sweep gut: sweep gutted stubs across agents, auto-reply, commands, cli (MBP audit) #2303) — src/agents/pi-extensions/context-pruning/pruner.test.ts (parent dir absent from fork; now explicit EXCLUDE-GUT)
• 1 rebranded — apps/macos/Tests/OpenClawIPCTests/… exists in fork as apps/macos/Tests/RemoteClawIPCTests/…

Full per-candidate verdicts: see HQ .tmp/systematic-validate-*/systematic-report.md and full-results.tsv.

Execution plan

Recommended order (can be tackled batch-cherry-pick style or one at a time):

1. Priority P1 — v2026.3.1 browser/output-atomic.ts + v2026.3.7 browser/session-tab-registry.ts (+test) — REJECTED as false positives (fork-gutted by 1aae7daa00 — Playwright removal, pre-B10). See § Rejected section.
2. Priority P2 — v2026.3.12 mattermost (reply-delivery.ts + test), v2026.3.22 msteams webhook-timeouts.ts, v2026.3.22 zalouser (zca-constants.ts + mocks) — channel-adapter fixes. 3 commits, ~5 files.
3. Priority P3 — Remaining 23 test files + .npmignore — coverage catch-up. 23 commits. Batch cherry-pickable when convenient.

Each gap is independently cherry-pickable with git cherry-pick -x {commit} (pattern established in #2402). Expect minor adaptations (rebrand transforms, fork-divergent import paths).

Prevention going forward

• INCLUDE Reconciliation Gate added to fork-sync SKILL § Phase D.3 — makes the silent-revert pattern a pre-commit invariant check.
• Reusable enforcement script at HQ scripts/reconciliation-gate.sh.
• Future sync batches will fail-fast on INCLUDE != committed_INCLUDE + reclassifications.

Notes

• This audit only covered fix: commits (higher signal). Non-fix commits (feat:, chore:, refactor:) in sync ranges may have additional lost files but are lower priority.
• The v2026.3.22 sync's deferred-extracts list (1076 files) is separately tracked in HQ (upstream/b10-deferred-extracts-list.txt). Most are expected-deferred (fork-divergent EXTRACT); this audit surfaces the subset that shouldn't have been.

Closes the "lost upstream fixes" question from #2362 triage.