activate/learn/amend lack --parent: recall Steps 2.5C/2.7/4 and learn cannot reach the parent vault from a child without an undocumented ENGRAM_SERVER per-call override

[toejough]

PROBLEM. vault-merged-recall gave show and show-chunk a --parent flag so an agent can resolve a from_parent item where it lives. The write-side commands the recall/learn procedures need next — activate (Step 2.7), amend (Step 2.5C), learn (Step 2.5C absent-write handoff, Step 4, and the learn skill itself) — have no --parent (internal/cli/activate.go, learn.go, amend.go: zero references at 1c741828), even though all three are already in the served command set (vault-serve-api: query, query-chunks, show, show-chunk, activate, learn, amend).

Consequence on a child. Every one of those writes targets the local vault by default. On a child with no local vault (the /app/repo sandbox today: ~/.local/share/engram/ holds only chunks/), activate on a parent note cannot find it, and a learn would mint a brand-new local vault that diverges from the parent — the opposite of what a child should do. The recall skill's own contract ("used notes must stay warm or the recency-competition mechanism breaks") is unmeetable from a child as written.

Workaround, verified 2026-08-30. Per-invocation server routing works: ENGRAM_SERVER="$ENGRAM_PARENT" engram show 824 resolved the parent note; the same override on activate/learn/amend reaches the served endpoints, and a served learn/amend lands as a pending offer for curate — which is the right semantics for a child contributing upward. But it cannot be set globally: Decision 5 makes ENGRAM_PARENT inert whenever ENGRAM_SERVER is set, so a node that exports ENGRAM_SERVER loses merged recall. It has to be a per-command trick, and neither recall/SKILL.md nor learn/SKILL.md mentions it — both are written as if the vault is always local (SKILL.md:26 "Do not pass --vault or --chunks-dir"; :207-216 bare engram activate --note …).

Proposal.

1. Add --parent to activate, learn, and amend with the same semantics show/show-chunk already have (route that one call to ENGRAM_PARENT's served endpoint; error if ENGRAM_PARENT is unset; inert under ENGRAM_SERVER). Served-write identity/offer rules (vault-serve-api, vault-offer-curation) apply unchanged.
2. Recall/learn skill text: on a merged node, items tagged from_parent: true are read with show --parent / show-chunk --parent and written back with activate --parent / amend --parent; new notes derived from parent candidates are written with learn --parent (landing as offers). This is a rung-2 skill edit — writing-skills TDD with a merged-mode fixture.

Related. #729; sibling issues filed the same day (zeroed budget block; explore displacing direct matches under --limit; parent clusters dropped). Note the spec's own rationale for an explicit flag over silent fallback (design Decision 8: bare Luhmann ids collide across independently-minted vaults) applies with more force to writes than to reads — an activate that silently hit a colliding local id would bump the wrong note with no error.

Siblings (filed 2026-08-30 from the same recall run): #743 (zeroed budget block), #744 (explore displaces direct matches under --limit), #745 (parent clusters dropped)

Addendum (same session): served activate hides partial failures from the client. RunActivate (internal/cli/activate.go:54-72) logs each miss via deps.LogWarning and only errors when every path fails (errActivateAllFailed); over the wire those warnings stay on the server's stderr, so a child calling ENGRAM_SERVER=… engram activate --note a --note b gets exit 0 even if b was skipped. Verified: a single bogus name returns POST /activate: activate: all note paths failed; a 17-note batch returned exit 0 with no output. Whatever shape --parent takes, the served response should carry the per-note skip list.

[toejough]

Superseded by local-first-parent-sync (main 242226b). There is no per-call --parent flag. Every environment now keeps its own local vault: learn, amend and resituate write locally and offer the note to ENGRAM_PARENT. engram activate on a parent note pulls it down as a pending local offer (design D8). ENGRAM_SERVER is now a hard error that points to ENGRAM_PARENT.