Beagle SCM

BRO-025: three-part O-link spells — context · verb · args

O click-spells encode only verb args; the context is inherited from wherever the pager happens to sit, so a board row cannot say "run this verb OVER THERE". Live incident 2026-07-14: the :work board's update minted get ///be for every row — authorityRepo hoisted ///be as the run scope and refreshed the LANDED journal/be mount; the row's worktree was never addressed, and the status shown after was the pager CONTEXT tree (//BRO-024), read by the user as the update result. Ruling: an O spell carries all three parts; the BRO-024 invite //WT/dir/: verb args IS the wire form; an absent context part means "here" (pager cwd).

Input

Context

• BE-043's :work board buttons must target the ROW's worktree; today all rows mint the byte-identical get ///be (views/work/work.js, uncommitted in work/BE-043); see BE-043.
• Evidence 2026-07-14: the click appended 26714CGTBr get ?#33295788 to journal/be/.be (a no-op refresh of the landed mount); work/BE-043 and work/BRO-024 untouched.
• authorityRepo (core/loop.js:189, BE-032) hoists a //-carrying arg0 of a MUTATION verb into the run scope — one arg0 cannot be both the scope and an operand; get.js:284-291 defers the explicit-//X-source vs injected-context split as a named follow-up.
• BRO-024 groundwork: context = process cwd (ONE-VAR), the invite is //WT/dir/: verb args, and replay already chdir's into the RECORDED view.call.context — the third part exists on the replay path.

Goals

• O token wire form = the invite line //WT/dir/: verb args; a context-less O token keeps meaning "here".
• The pager drives an O spell IN the spell's own context (the BRO-024 chdir/serialize/restore path), never pre-merging context into args.
• Spell-driven verb runs see (context, RAW args); an explicit //X arg is a plain operand for the verb, not a rescope.
• :work board rows mint //KEY/: get ///be — update = FF the KEY tree to the landed mount's commit, local changes preserved (weave only where dirty).
• U nav tokens unaffected (already full addresses).

Constraints

• Verbs-resolve-args law intact: pager arg-blind, resolve(context, base, rel) at verb entry; the cd is the backstop (BRO-024).
• CLI-typed jab <verb> //X … keeps the BE-032 arg0-scope convention until ruled otherwise; the change here covers SPELL-driven runs.
• get's source-form (get ///be = FF to that tree's commit) is GET-side; split into its own GET ticket if it grows past the seed.

WIP

Design decisions

• (gritzko, 2026-07-14) O-links are three-part: context, verb, arguments.
• (impl, 2026-07-14) the wire codec is SPELL.parseOspell/mintOspell (shared/spell.js): a leading // ARMS the split at the FIRST : , so a : inside a quoted arg (post 'a: b') is safe; a context-less token (verb args) parses to context "" = "here" (the pager context).
• (impl) the pager drives an O click through _actSpell(spell, ctxOverride) / _runSpell(spell, ctxOverride) — the CONTEXT comes from the parsed O token, threaded to driveSpell (opts2.context), NEVER pre-merged into an arg; _actSpell then _refreshes the pager's CURRENT view IN PLACE, so the pager's own context is left UNMOVED (the "restore"); a throwing drive is caught, context still stands. There is no literal io.chdir on the reentry path — the context scopes the run via loop.contextRepo (be.repo/ctxDir), the mechanism replay already uses.
• (impl) loop.cli scopes CONTEXT-FIRST: a threaded context (contextRepo) wins and keeps args RAW — NO authorityRepo arg0 hoist on the spell-driven path; the BE-032 arg0-authority hoist stays for CLI-typed jab <verb> //X … (no threaded context). uri011 legs stay green (CLI convention intact).
• (impl) //KEY/: rm . per gritzko's spec — the context //KEY fully targets the row, . is the whole-worktree operand (rm verb not landed in this tree; token-text only asserted).
• (impl) get.sourceGet: a scheme-less //X////path operand routes through the GET-038 local-source machinery (parseRemote("file:"+wtdir) → resolveLocalSource → seedLocal → fanoutWholeTree) — a clean FF fast-path, a dirty baselined file weave-merges (the same leaf). This reaches get ONLY on the spell-driven RAW-arg path; a CLI jab get //X still hoists arg0 to the run SCOPE (whole-tree FF), unchanged.

TODOs

• spell codec: parse/mint the //WT/dir/: verb args O token form (shared/spell.js parseOspell/mintOspell); absent context = "" ("here"); round-trip + quoted-: safety proven (test/bro/ospell, work/view)
• pager: drive an O spell in ITS context (_actSpell/_runSpell ctxOverride → driveSpell context), pager context UNMOVED on success / mutation / throw (test/bro/ospell)
• work.js: mint per-row //KEY/: get ///<mount>, //KEY/: post '<title>', //KEY/: rm . via SPELL.mintOspell (test/work/view updated + green)
• loop: spell-driven runs pass (context, RAW args) — contextRepo wins first, no arg0 authority hoist on that path (core/loop.js _cli; uri011 + status/put/post/delete/patch all green)
• get seed: an explicit //X////path operand = SOURCE worktree (get.js isSourceWorktree/sourceGet, GET-038 resolveLocalSource), dest = the run context; FF fast path, weave-leaf when dirty
• tests: board update advances the ROW tree — not the landed mount, not the pager context tree (test/get/source-worktree, END-TO-END through the real driveSpell: cwd SITS IN a LAND tree, context //DEST, only DEST advances)

Blockers and bummers

• cross-STORE source-get (dest store lacks the source tip's objects) needs an INGEST — deferred as a GET follow-up; the hive's shared-store cells (the :work board case) need none (source k materialises directly, dest .be tip appended). Filed nowhere yet — raise a GET ticket if a cross-store update is ever wanted.
• Nav.mkd's three-part-O-link spec rides BRO-024's still-open "rewrite the Nav context sections" TODO (Nav.mkd is flagged STALE); not expanded here to avoid double-writing the pending rewrite.
• pre-existing baseline reds unrelated to BRO-025 (each fails identically in the landed tree / is documented): get/oddname (the #-framing pair), sub/wtsrc (the runtime .be/?/ query-slash strip, BRO-024 _restoreQuerySlash shim), bro/color + bro/view (stale goldens), bro/vim (fails even with the BRO-025 pager edits reverted).

Outcome

• iter 1 (all six TODOs) implemented in work/BE-043 (UNCOMMITTED, per order): shared/spell.js (+parseOspell/mintOspell), views/bro/pager.js (mouse O-parse + actSpell/_runSpell ctxOverride), views/work/work.js (three-part mint), core/loop.js (context-first RAW-arg scoping), verbs/get/get.js (isSourceWorktree/sourceGet).
• new tests: test/bro/ospell (15 headless pager checks), test/get/source-worktree (real shared-store FF, cwd-vs-context control); test/work/view updated to the three-part O tokens.
• local suite 2026-07-14: 144 PASS / 5 FAIL over test/// (network dirs excluded); all 5 are the pre-existing baseline reds above. Command: for run in test/*/*/run.sh; do timeout 90 sh "$run"; done with JABC=/bin/jab, BIN=/bin, ASAN_OPTIONS=detect_leaks=0, TMP=/tmp/claude-1000.