B5: one create path — behavioral quad activation, backfill, menu collapse, legacy dialog deletion

[wesbillman]

B5 (Bundle 2): one create path — behavioral quad activation, backfill, menu collapse, legacy dialog deletion

What this PR does

Completes Phase 2 Bundle 2 of the unified agent model: every agent is created through ONE flow (the definition dialog), the NIP-AP behavioral quad (respond_to / allowlist / mcp_toolsets / parallelism) is fully live end-to-end, and the legacy standalone-instance CreateAgentDialog is deleted.

13 commits, each independently reviewed in-channel before landing:

1. 57927b5ef — activate NIP-AP behavioral defaults: definitions carry the quad, applied at instance mint (resolve_mint_behavioral_defaults; input wins, else definition quad, else defaults; loud failure on invalid definition values).
2. f7472a212 — apply inbound quad edits on the persona-match branch.
3. 0e61f2526 — backfill standalone agents into definitions: manufactures a definition per pre-existing standalone agent (slug = agent pubkey), idempotent, .bak backup, hash-quiet (snapshots record's own values so spawn_config_hash and the drift badge stay quiet), system_prompt present-even-if-empty (old readers hard-fail on absent prompt).
4. cc1547761 — share the spawn-effective prompt filter, exempt team packs.
5. 4dfac6caa + 685cfe22a — sync-flush ordering: tombstones publish before replacements, and replacements defer behind a failed tombstone (race found during review).
6. 1dae8a9a3 — file-size split (persona_events tests + pending-retention helpers).
7. 75cd8d771 — menu collapse: "Custom agent" menu entry removed; the app-wide create-agent shortcut routes to the unified definition flow with the start-after-create toggle.
8. 60555c41b — provider/mesh re-home: "Where to run" (local / discovered provider / relay mesh) lives in the unified create dialog; mesh preflight runs BEFORE the mint (fused in mintDefinitionWithPreflight, reordering structurally impossible); stale-intent guards on both doors.
9. 18c8f2779 — split persona request types into types/requests.rs (pure move).
10. 6daa7e8d0 — definition-quad dialog fields: who-can-talk (respond-to + allowlist), toolsets, parallelism editable on agent definitions; apply_persona_behavior fuses validation unskippably into create_persona AND update_persona; crash-loop guard (empty allowlist blocks submit) in create and edit; hash-quiet dirty-detection (untouched quad on edit submits no behavior group, so renames never rewrite published quad bytes).
11. 2abbf4aac — delete CreateAgentDialog + Sections (−1,290 lines), collapse the instance create intent, migrate the e2e suite with mock-bridge mint-parity, fix the edit full-clear resurrect.

Disclosures (argued decisions, in commit bodies)

• One-time republish wave: activating the quad + backfill triggers a second fat→slim republish wave for definition-linked records (same shape as feat(relay,desktop): canonicalize agent definitions on kind:30175 (Phase 2) #1655's); idempotence covered by test rows.
• Ordering window accepted: a slim 30177 can land before its 30175 on an old device; self-heals when the definition lands — ordering machinery failed the minimal-fix bar.
• Relay NIP-09 looseness: relay accepts deletes loosely; tombstone-before-replacement ordering in the sync flush is the client-side mitigation.
• Create-time capability regressions: relayUrl / custom commands / turnTimeout are no longer settable at create (they remain editable on the instance after mint). Env-vars semantic: definition env layered under instance overrides — creates no longer seed pseudo-overrides.
• Start-toggle-off = no instance: definition-only create mints nothing; "where to run" state is discarded (stale-intent guard, tested both doors).
• Anthropic-without-model now blocks submit (legacy allowed save + post-spawn nudge); the unified dialog's pre-existing gate is stricter. Readiness shot 02 asserts the stricter gate.
• Mesh honest copy: the mesh preset never overwrites definition fields — only the minted instance carries mesh commands/env; rendered as user-facing copy under the mesh section.
• Old-device republish ping-pong + Some("") badge flip: one-time, self-terminating; discussed in-channel.

Testing

• Rust: 1115/1115 (cargo test --lib, post-rebase onto e0f76b0), clippy clean (2 pre-existing warnings), both fmt passes.
• Frontend: 2245/2245 unit, tsc, biome, all three check scripts.
• E2E: full suite verified locally against an isolated relay before push — 536/540, the 3 failures bisect-proven pre-existing on main or flaky-cleared (migrated specs keep their END assertions — the smoke parallelism assertion can only pass through the real definition-quad mint fallback, mirrored honestly in the mock bridge).
• Every commit reviewed by Pinky against pre-agreed pins (PLANS/B5_REVIEW_GATES.md), including three review-found bugs fixed in-flight (tombstone race, edit full-clear resurrect, migrated-spec persona-card gap).

12. 18961f24d — sweep dialog-deletion orphans (AgentProviderField/AgentModelField, zero importers) + fix the mesh restart spec for the persona-card grouping the definition flow produces (reveal-dialog dismissal reordered before card assertions; END assertions untouched).

[wpfleger96]

🤖 Consolidated team review — Paul (orchestrator), Duncan (implementer-lens), Thufir (independent analyst), Alia (copy + mechanical verification).

All four reviewers independently examined the PR at head 18961f24d in detached worktrees. Duncan and Alia: Approve. Thufir: Request changes (one production correctness gap — provider payload drops mcp_toolsets). Paul's initial assessment missed the provider payload issue; Thufir caught it, verified at source.

Finding not attached inline (file not in diff):

IMPORTANT — Edit-path empty-allowlist guard untested (convergent: Paul, Duncan, Thufir) — personaRuntimeModel.ts:258: computeEditAgentFormValidity rejects respondTo === "allowlist" with an empty allowlist, but the edit-path tests (agentInstanceEditPinning.test.mjs:165) use respondTo: "mentions", so the guard is never exercised. The create-path guard (personaBehaviorDraftValid) has a direct test. Fix: add one false case with respondTo: "allowlist" + respondToAllowlistLength: 0 → false, and one true case with length > 0 → true.

See inline comments for all other findings, each tagged with provenance and severity.

[wpfleger96]

🤖 IMPORTANT — Provider deploy payload drops mcp_toolsets (Thufir-sourced, Paul source-verified)

build_deploy_payload (lines 367–399) serializes parallelism, respond_to, and respond_to_allowlist into the provider agent_json, but omits record.mcp_toolsets. The local spawn path correctly sets BUZZ_TOOLSETS from the record at runtime.rs:1768–1772. All three provider paths (create at line 961, start at line 1070, redeploy) funnel through build_deploy_payload, so provider-backed agents silently lose inherited or explicit toolset settings.

Fix: Add "mcp_toolsets": &record.mcp_toolsets to the serde_json::json! block in build_deploy_payload, and add a regression test for provider-create payload completeness.

---

should-fix — Parallelism input validation at wrong layer + misleading doc (Paul-raised; Duncan and Thufir independently corrected the framing)

resolve_mint_behavioral_defaults at types.rs:932 passes explicit input parallelism through without range validation, while the definition branch validates loudly. The doc comment on MintBehavioralDefaults.parallelism ("Validated (1..=32) when present") is false for the input branch.

However, call-site validation at agents.rs:525–529 guards the full create path before this function is reached — so the behavior is not broken end-to-end. This is a wrong-layer defense + inaccurate doc comment, not a live behavior defect.

Fix: Either add range validation inside the function's input branch, or update the doc comment to state the caller is responsible for validating input parallelism.

[wpfleger96]

🤖 IMPORTANT — Mock-bridge mint-parity missing mcp_toolsets (convergent: Paul, Duncan, Thufir)

handleCreateManagedAgent (~line 6587) resolves respondTo, respondToAllowlist, and parallelism from the linked definition, but has no mcp_toolsets resolution at all. An e2e test asserting toolset inheritance would pass regardless of correctness.

Fix: Add mcpToolsets to the mock input, add mcp_toolsets to RawManagedAgent/clone output (~lines 440–475, 1027–1063), and resolve as input ?? linkedPersona?.mcp_toolsets ?? null.

[wpfleger96]

🤖 MINOR — User-facing copy uses internal jargon (Alia)

Line 72–73: "Create an agent profile and start its managed agent instance" → "Create an agent and start it immediately."
Line 50: "without starting an instance" → "without starting it."

[wpfleger96]

🤖 MINOR — Grammatically incomplete toggle label (Alia)

Line 135: "Start agent after create" → "Start agent after creating" or "Start agent after creation."

[wpfleger96]

🤖 MINOR — Inconsistent branding + missing article + unused prop (Alia + Paul)

1. Inconsistent branding: "Run on relay mesh" (line 122) vs "Relay compute" (line 67) vs "Relay mesh" (line 170). Pick one canonical term.
2. Line 90: Missing article — "Selected relay mesh target…" → "The selected relay mesh target…"
3. Line 45: modelId: string prop declared but never destructured or read in the component body. (Alia CONFIRMED)

[wpfleger96]

🤖 MINOR — Parallelism error string inconsistency (Alia)

Line 940: Error string prefixes with "definition" ("definition parallelism {count} is out of range…"). The corresponding string in types/requests.rs:55 does not. Both can surface to users. Align one way or the other.

[wpfleger96]

🤖 MINOR — Button label is a state, not an action (Alia)

Line 27: deselectAction: "Selected" — the paired aria label (line 52) correctly says "Deselect". The visible button text should match the action: "Deselect" or "Remove".

[wpfleger96]

🤖 MINOR — Two nits (Paul + Alia)

1. Line 11: PERSONA_LABEL_OPTIONAL_CLASS redefined byte-identically — already exported from personaDialogPickers.tsx:8, and this file imports two other constants from that module. Import instead of redeclaring. (Alia CONFIRMED)
2. Line 90: Parallelism hint "Concurrent turns each started instance handles (1-32)" → "How many conversations each running instance handles at once (1–32)." (Alia)

[wpfleger96]

🤖 MINOR — Dead export (Paul, Alia CONFIRMED)

Line 352: openCreate() defined and returned at line 449, but has zero call sites in desktop/src. B3-migration leftover — delete or add TODO(B3).

[wpfleger96]

🤖 MINOR — No partial-input quad test (convergent: Paul, Duncan, Thufir)

Lines 555–635: mint_explicit_input_wins_over_definition passes all four fields at once. No test covers field-by-field independence (e.g., explicit mcp_toolsets + inherited respond_to). Add one mixed-input test to pin the field-by-field contract.

[wpfleger96]

🤖 MINOR — Slug-collision skip is silent and permanent (convergent: all reviewers)

Lines 84–89: An agent whose pubkey collides with an existing definition's slug is skipped with eprintln! and stays persona_id: None forever. Worth a surfaced warning or a comment documenting the recovery path (delete the colliding definition).

[wpfleger96]

🤖 IMPORTANT (pre-existing, elevated by this PR) — Two security boundary gaps (convergent: Paul, Duncan, Thufir)

1. No sig verification — Line 386: nostr::Event::from_json skips sig checks. This PR routes the behavioral quad through this path, so a forged event could now change who may command an agent.
2. Tombstone owner unchecked — Line 491: parse_deletion_coordinate discards _owner; a forged kind:5 could delete another owner's record.

Fix (follow-up PR): Add event.verify() after parse + require tombstone coordinate owner to match event.pubkey.

[wpfleger96]

🤖 MINOR (pre-existing) (Paul, Duncan, Thufir, Alia confirmed)

1. Lines 234–280: Flush loop opens the retention DB ~3× per pending row. B5's republish wave makes this O(N agents) on first post-upgrade boot. Hold one connection across the sweep.
2. Line 296: persona_content_hash uses unwrap_or_default() on serde_json::to_vec — silently hashes empty bytes on failure. expect() would fail visibly.

[wpfleger96]

🤖 MINOR (pre-existing) (Paul, Alia CONFIRMED)

Lines 1367–1375: spawn_config_hash recomputed per running agent on every list_managed_agents poll (~30s). This PR added quad fields to the hash input. Cache-on-stamp candidate someday.

[wesbillman]

All review findings addressed in acbe4d3a0:

IMPORTANT:

• Provider payload mcp_toolsets — added; build_deploy_payload split into agents_deploy.rs with a pure deploy_payload_json half and a payload-completeness regression test (deploy_payload_carries_the_full_behavioral_quad), so a missing behavioral field now fails a test instead of silently stripping provider agents.
• Mock-bridge mint-parity mcp_toolsets — resolved as input ?? linkedPersona?.mcp_toolsets ?? null, threaded through RawManagedAgent + both clone paths.
• Edit-path allowlist guard test — two-case row added (editValidity_allowlistWithEmptyList_blocksSave): empty list blocks Save, non-empty allows.

should-fix: input-branch parallelism now validated inside resolve_mint_behavioral_defaults (the doc contract is now unskippable, not call-site-delegated), with distinct input-vs-definition error strings + tests (mint_rejects_out_of_range_input_parallelism, mint_resolves_each_quad_field_independently).

Minors: all copy fixes applied (create/start descriptions, "Start agent after creation", canonical "relay mesh" + articles, Deselect action label, plain-words parallelism hint); dead modelId prop, dead openCreate export, and the duplicated class constant removed; backfill slug-collision skip now documents retry semantics + recovery path in comment and warning text.

Deferred as the review itself framed them (tracked in the workstream backlog): the two pre-existing security gaps (inbound sig verification, tombstone owner check — follow-up PR) and the pre-existing perf notes.

Also riding: a small owner-requested UX addition — saving an edit on a stopped/failing instance now offers a "Start now" action toast, since the auto-restart policy deliberately never fires for broken agents. Policy gates unchanged.

Gates on the new head: Rust 1118/1118, frontend 2246/2246, tsc/biome/checks/fmt clean, edit-agent + readiness e2e ✓.