Beagle SCM

BEE-003: the repo is an axis of the target — pager and http serve every local repo

A page in one repo links a page in another — the wiki links tickets, tickets link code — but in bee the repo is PROCESS STATE, not part of the address: http.js opens the cwd's repo ONCE and chdirs to its root, and the door resolves every partial at io.cwd()'s repo. So a cross-repo ref renders as a link into the serving tree and 404s, or resolves to a same-named local file, and neither the pager nor the server can switch repos at all. BEE-002 makes the BACKLINK direction cross the boundary; this makes the repo an axis of the TARGET, which is what lets both apps serve every registered repo.

Input

Context

http.js takes the door's openPath/seatOf/statOf and serves paths of one repo; render/html.js turns a ref token into an href; index/resolve.js resolves a partial path against ONE tip tree.

• http.js:405-407 opens io.cwd()'s repo once and CHDIRS to its root, so every URL is relative to that one tree — global state no second repo can share.
• door.js:94 resolves a partial at io.cwd()'s repo too, so the pager inherits the same single-repo anchor and cannot switch trees.
• A ref is whatever hook.js fTokens fuses (the ONE recognizer) — the resolver, not the lexer, is what lacks a repo axis.
• Ref spellings in the corpus: ///bee/index/index.js (authority), /todo/BEE/BEE-001.mkd (root-absolute, another repo's root), ../wiki/Plan.mkd (relative, climbing out of the repo root), plus bare ticket codes.
• BEE-001 landed $HOME/.config/bee/repos — the list a resolver can mount.

Goals

The pager and bee http both serve every registered repo, and a cross-repo ref opens the right page in the right repo instead of a same-named local file.

• The REPO is part of the target the door opens, not the process's cwd; the http chdir retires.
• Every registered repo is reachable over the served URL space, one prefix per repo.
• Following a cross-repo link in the PAGER switches trees in the same session, back/forward included.
• ///name and root-absolute refs resolve through the registry; a relative ref that climbs out of its repo root resolves to a sibling repo or is refused in words.
• The plain/ansi renderers keep working unchanged — this is a resolution change, not a hunk change.

Constraints

• Read-only, as LITE-034 ruled for bee http; no verb gains a write.
• No new runtime binding; fTokens stays the one recognizer, resolve.js the one resolver.
• Serving a repo must not require indexing it: a registered repo with no lane still serves its files.
• Repro first, as a test under test/http/.

WIP

Design decisions

1. RULED: the repo is a field of the TARGET, carried in an AMBIENT context per run / per request / per view — {repo, path, anchor}, be's be.repo (shared/ambient.js) in bee's shape. The process cwd stops being the repo and the http.js chdir retires. The pager's back stack holds (repo, path, anchor) triples, so a cross-tree hop pops back cleanly.
2. RULED: the URL CARRIES the repo — /<repo-name>/<path> — and the registry is the mount table, so a request resolves from the URL alone and the server keeps no current-repo state. A bare repo-less URL 301s to the prefixed form; it does NOT keep working silently, or old links never converge on the addressable shape.
3. RULED: a partial ref resolves in ONE fixed order — the DIR of the file being read, then the current repo at HEAD, then every registered repo (recursing submodules, BEE-006) — and a miss is a WORDED refusal that names what was searched ("no registered repo holds abc/TCP.c"), since the useful next move is bee install on the repo that owns it.
4. RULED: several hits land in the CHOOSER hunk the resolver already builds for ambiguous partials, each row naming its repo — deduped by blob sha and mount identity, so a submodule file reached through the parent AND through its own lane is ONE row, not two.
5. RULED: a submodule file is canonically addressed THROUGH the parent — /quickjab/dog/abc/TCP.c — and served at the commit the parent's gitlink PINS, not at the sub's own HEAD (they differ whenever the sub is ahead). The sub's own mount is the same bytes at another commit: it redirects, or is marked as the origin, but both are not live addresses for one file.
6. OPEN: repo NAMES collide already (bee and bee2, plus worktrees of one repo). The registry basename is the default name; the disambiguator when two entries share one is unruled.
7. RULED (was candidate 2): a relative ref that climbs out of its repo root is resolved by the resulting absolute path against the registry; no registered repo contains it → the link renders dead, in plain words.
8. RULED: the per-repo lookup is FSEG (LITE-011, index/resolve.js), never the REV lane. path_hl hashes the WHOLE path, so no REV row is findable from a suffix; FSEG is keyed fn_hl:40|prnt_hl:20 — the shape a partial actually gives — and its ancestor segment hashes prune the tree descent to a line, recovering real TEXT at the bottom and answering ambiguity with several paths. Fanning that scan over the registry is a prefix scan plus a pruned descent per repo, no blob reads. REV-PARS stays the weave's DAG edges (BEE-005), a different question.

TODOs

• repro test: a cross-repo ref renders as a link that 404s or mis-points
• the ambient {repo, path, anchor} context; openRepo(io.cwd()) and the http chdir retire
• the resolution ORDER of ruling 3, dir-of-the-file leg included (missing today)
• the FSEG scan fanned over the registry (ruling 8), one repo ctx per hit
• registry-backed resolution for ///name, root-absolute and climbing refs
• /<repo-name>/<path> URLs off the registry; bare URLs 301 to them
• the chooser rows name their repo and dedup by blob sha + mount identity
• a submodule file resolves at the parent's PINNED commit (ruling 5)
• the pager follows a cross-repo link into the other tree, back/forward intact
• a ref naming no registered repo refuses in words, naming what was searched
• docs: README.mkd http lines, INDEXES.mkd if the lane is touched

Blockers and bummers

• Evidence for ruling 4: abc/TCP.c exists in SEVEN sibling repos here (c-rdx, libdog, librdx, dogs, strictmark, causal, ast-rev-ctrl) — the ambiguity is the normal case, not an edge one.
• Only registration keeps the chooser small: ~/.config/bee/repos holds six TOP-LEVEL repos and no subs, so of those seven only the one under registered quickjab (dog/abc/TCP.c) is reachable at all.
• /home/gritzko/src/bee/bee is a SYMLINK to the repo root itself, so "reading in bee/bee/" never leaves the bee repo — a resolver must not take a symlinked path for another tree.
• Today door.js seatOf skips its own worktree-scan fallback whenever a repo IS found (resolvePartial answers [], not null), so a cross-repo ref dies as a bare "cannot open" with no search at all.

Outcome

(open)