fleet-memory/FORK.md

RCLL as a fork of Hindsight

RCLL is a fork of vectorize-io/hindsight (MIT, Copyright (c) 2025 Vectorize AI). This file records what we changed, where we branched from, and how to take the next upstream release — the three things the repository could not answer before, and the reason updating had become archaeology instead of a rebase.

What was wrong

The code arrived here on 2026-06-27 as a single commit of 1716 files and 558,283 lines, with no upstream remote and no common ancestor. Five later commits carried branding only. So git log could tell you that a rebrand happened and nothing at all about the engine changes underneath it — and git merge upstream/main was not merely painful, it was impossible: unrelated histories.

Where we actually branched

Not from a release tag. The snapshot was matched against upstream commit by commit, minimising the number of differing files:

candidate	files differing from our tree
v0.5.0 (2026-04-08)	146
7b2263ba (2026-04-13)	34
d054b884 (2026-04-10)	32

Base = d054b884, "fix: add PEP 561 py.typed marker to all Python packages (#973)", 2026-04-10 — twelve commits past v0.5.0, which is why an earlier reading of the delta against v0.5.0 looked like 146 changed files and 13,383 inserted lines. Most of that was upstream's own work between the tag and our branch point, not ours. Our real patch is 32 files.

The series

rcll/main replays that patch as nine commits on top of d054b884:

commit	what
chore: drop committed session keys and .DS_Store	two .sesskey files and a .DS_Store that upstream had committed
ci: replace upstream release pipeline with npm Trusted Publishing	our publish workflow; upstream's four release/test workflows removed
ops: standalone compose and env template	docker-compose.rcll.yml, .env.example, standalone entrypoint
feat(rooms)	per-agent rooms, halls, durability layers — ADR-145. The one thing in this repo that is ours and has no upstream counterpart
feat(closets)	compression of memories by room+hall — ADR-145 phase 3
feat(tunnels)	cross-bank memory bridges — ADR-145 phase 4
fix(engine)	sync migration URL fallback, protected tables, small engine fixes
feat(mcp)	the RCLL MCP server and server.json registry manifest
brand	README.md, RCLL.md

How the split was verified

The series was not hand-sorted. scripts/rcll-split.py decomposes the base→snapshot diff to line granularity — hunks are too coarse, because the two largest are single appended blocks containing more than one feature — then classifies every edit and materialises each stage by content.

Two checks, both mechanical:

• Identity. The tree at the last commit is byte-identical to the snapshot tree (git diff --quiet <series-head> <snapshot-tree>). A split that lost work could not pass this.
• Each stage compiles. Every intermediate commit is checked with py_compile across hindsight_api/. This caught a real bug in the splitter: a first version deleted base lines for not-yet-staged edits, which still satisfied the identity check — because the final stage enables every group — while corrupting all eight commits before it.

Known imperfection, stated rather than hidden

Ten lines of tunnel code sit in the feat(closets) commit: one DELETE /tunnels/{id} endpoint pair, plus the shared _PROTECTED_TABLES entry that adds "tunnels" and "closets" on the same line. Rooms, closets and tunnels are one ADR and genuinely share plumbing; this is where the automatic boundary stops being sharp. Everything else separates cleanly (rooms contains zero tunnel or closet lines; tunnels contains zero closet lines).

Distance to upstream

As of 2026-08-23, upstream main is 3295716c and the newest tag is v0.9.1 (2026-08-14). From our base that is ~1600 first-parent commits, and the four files our room/hall work touches most are also the four upstream rewrote most (memory_engine.py, http.py, orchestrator.py, retrieval.py). Upstream has no room/hall/layer concept at all — checked by grep against main — so there is nothing to inherit and nothing to drop.

Our alembic line adds exactly one migration upstream does not have, aa1_add_room_hall_to_memory_units. Production currently has two alembic heads (aa1_room_hall and upstream's h3i4j5k6l7m8); any upgrade needs an authored merge revision first.

Taking the next upstream release

git fetch upstream --tags
git rebase --onto v0.9.1 d054b884 rcll/main    # nine commits, not one blob

Conflicts land inside the feature commit that owns them, which is the entire point of the series.

Do not run migrations as part of this. On Ring 0 the memory schema hindsight_v2 lives inside the CRM database, so an upstream migration is DDL against a combat master, and 16 of the ~50 new ones are destructive (drop_memory_units_access_count, drop_entity_memory_links, split_history_into_own_tables, …). HINDSIGHT_API_RUN_MIGRATIONS_ON_STARTUP stays false there. Rehearse on a copy of the schema, never on the master.

Remotes

origin    https://github.com/holetron-lab/fleet-memory.git
upstream  https://github.com/vectorize-io/hindsight.git   (push URL deliberately
                                                           poisoned — nothing on
                                                           this box can push to
                                                           vectorize-io)

What is deliberately NOT done here

main is untouched. Making main descend from upstream requires a force-push, and this repository is one decision away from being made public with an already published server.json — a rewritten history is a one-way door. The series lives on its own branch; repointing main is the owner's call.

Dependency updates: why there is no dependabot

.github/dependabot.yml was removed on 2026-08-28. It was not failing — it was running in a loop that could never land.

GitHub is a push mirror of the Forgejo canon, not a peer. The mirror sync force-matches refs, so a branch dependabot opens on GitHub is deleted within the hour and its pull request closes itself. Two runs proved it (bump actions/checkout 4 → 7 and bump actions/setup-node 4 → 7, both opened 2026-08-24, both auto-closed). A public repository whose only pull requests are self-closed bot PRs reads as abandoned, which is the opposite of what the bot is for.

The bot's one useful output was taken by hand: .github/workflows/publish.yml now pins actions/checkout@v7 and actions/setup-node@v7, and that upgrade shipped in the 0.1.0 release.

Action versions are therefore bumped on the canon, by hand, when the workflow is touched. If GitHub ever becomes the canon rather than a mirror, restore the file — the loop closes only because pushes are one-way.