Fork feature map and upstream-overlap report in the rebase routine

[lukemaj]

Outcome

A feature map of everything the fork adds on top of upstream T3 Code, used by the rebase routine to spot upstream changes that duplicate, conflict with or make a fork feature obsolete, so an upstream absorption never leaves two versions of the same capability.

Why (user, 2026-09-25)

docs/fork.md and scripts/fork-upstream-edits.txt list edited upstream files, not features. Example risk: we add auto-resume after a usage limit, upstream later adds it elsewhere, and after the rebase both exist without anyone noticing.

Acceptance criteria

• docs/fork-features.md: one entry per fork feature with id, one-line purpose, the Issue and PR, new files, edited upstream files, and "watch for upstream" keywords (for example: parentThreadId, child thread, subagent, usage limit, resume, sidebar section, branding).
• Every file in scripts/fork-upstream-edits.txt belongs to exactly one feature id; scripts/fork-check.sh fails when a fork-edited file has no feature.
• The rebase routine (scripts/fork-rebase.sh and docs/fork.md "Routine") produces an overlap report before absorbing: the upstream commits since the last absorption whose titles, touched files or diffs match a feature's files or keywords, grouped by feature. Each match needs a recorded decision in the absorption PR: keep ours, adopt upstream and delete ours, or merge both.
• Backfilled for every existing fork feature (branding feat(desktop): brand the fork build as Chromeria #12/feat(web): show Chromeria as the product name in the web UI #14/test(brand): expect the Chromeria web icons in production builds #13, child threads feat(threads): production child threads in Agents panel plus threads MCP toolkit (#8) #10, thread scope Threads toolkit: read and message any thread in the project (scope: children | project) #15, Agents panel Agents panel: Prism spawns section, collapsed child tree, parent breadcrumb #17, fork maintenance docs(fork): Chromeria install, rebase routine and GitHub-hosted CI #9, direction files, Prism toolkit Prism in Chromeria: MCP toolkit, spawn_thread(role), role settings schema, provider snapshot endpoint #19 once merged).

Blockers

Start after #19 (stream A) merges, because both edit docs/fork.md, the edit list and fork-check.

Objective contribution

Fork Objective: keep the fork a thin, rebasable layer on upstream.

[lukemaj]

Implementation starts from origin/main at 1ab79b5 in the exclusive chore/20-feature-map worktree. The wanted result is a complete feature inventory and pre-absorption overlap review, supporting the thin, rebasable fork Objective. Smallest solution: machine-readable entries within docs/fork-features.md, one dependency-free reader used by fork-check and fork-rebase, and real Git fixture tests. No separate registry, dependency, upstream absorption, or app/runtime changes. Shared files get one primary ownership assignment with additional feature references for overlap detection. Proof: fork-check, fork maintenance tests, typecheck, lint, fmt, independent review, and fork CI before the authorized squash merge.

[lukemaj]

Candidate: 446f35d35ff3bdbfdad01680bcf05736281b85b8, base: 1ab79b54b28d299d53aa9ae4c24de0474a804ca2.

All local checks passed on that exact candidate:

• bash scripts/fork-check.sh
• pnpm exec vp test run scripts/fork-maintenance.test.ts (12 tests)
• pnpm exec vp run -r typecheck
• pnpm exec vp lint --quiet
• pnpm exec vp fmt --check
• pnpm exec vp run knip:check

Independent review: Luna max in OpenCode, session ses_f26146be8ffe0K187csLZqFCgF, Model Router policy durable-runner-policy-v2 2.8.0. Native gpt-5.6-luna review route was unavailable in this Codex session, so the policy's luna-go-review fallback was used. Final verdict: No Blockers on the exact candidate. The initial stale-guidance finding was accepted and fixed. Initial CI's Knip unused-file finding was fixed by registering the shell-invoked helper as an entry point, with matching feature ownership.

Fork Stack Model passed. Remaining upstream CI is queued. The entire fork-new file set since b2b43bef7 is represented in the map, and every allowlist path has exactly one primary owner.

Website/SEO impact: none, this changes maintainer tooling and documentation only. No version transition: this fork has no version policy and adopting one is outside Issue #20. No actual upstream absorption, application run, installation, or user-data access was performed. User-authorized merge will use the single PR #24 once all gates are green.

[lukemaj]

Delivered in PR #24, squash-merged into main as 13b7a4ffdebf7d34dfda103874f62439bf3c0e0a at 2026-09-25 19:12:17 UTC. Live origin/main was verified at that commit, Issue #20 is closed, and the remote task branch is deleted.

All local proof and independent review passed candidate 446f35d35ff3bdbfdad01680bcf05736281b85b8. Full CI passed; fork CI passed. No remaining findings.

The feature map covers: direction files (#5), glossary (#11), Chromeria branding (#12/#13/#14), child threads (#10), fork maintenance and CI (#9), thread scope (#15/#16), Agents panel (#17/#18), Prism toolkit and role kits (#19/#22), Prism settings (#21/#23). It validates one primary owner per allowlisted path and drives the pre-absorption report of matching upstream commits, with per-match decisions required in the absorption PR.

No upstream absorption, direct push to main, app/runtime installation, or user-data access was performed. No version-policy adoption was included. Original checkout's untracked direction files were left untouched.

[lukemaj]

Finalization complete: removed the clean task-owned worktree and local chore/20-feature-map branch after verifying merge and remote branch deletion. Other worktrees and the running Chromeria were untouched. Agent Observer task t3code-20 reports complete coverage with no missing or conflicting bindings; private export retained in its local state directory.