snapshot: persistent-memory — v0.2.1 layout fix + §5 sequencing decision + Worldtree-stall diagnostic

Captures three things accumulated since the v0.2.0 snapshot in 3b9c610:

1. v0.2.1 layout fix (c85f6bd) — Recent decision documenting the
   dock-anchored TUI chrome that fixed Input bouncing. Operator-verified
   "a lot better" interactively. Going-forward principle: TUI-layout
   patches ship + operator verifies (TTY is the load-bearing test
   surface; respx/Pilot can't catch screen-relative positioning bugs).

2. §5 sequencing decision — collapsible Thinking pane + Debug pane
   proposals fold into design-brief §5's TabbedContent column rather
   than ship as inline-Collapsibles first. Do issue #8 (startup agent
   picker) before §5. Avoids the build-inline-then-rebuild waste.

3. Worldtree-stall diagnostic shorthand — "2-events-then-silence"
   = upstream LLM-provider connection wedged, not ratatoskr.
   worldtree-dev confirmed via code-level walk-through (althing
   01KSBKTG096Q…). Future-self defense against bisecting ratatoskr
   code when this shape appears.

Also refreshes the in-flight section: v0.2.0 + v0.2.1 shipped + tagged;
lofn smoke now unblocked on auth (still pending operator); next planned
feature is issue #8 (startup agent picker), then §5 side-panes.

No version bump per CLAUDE.md SemVer etiquette (memory-snapshot
commits skip).

persistent-memory.md

@@ -32,77 +32,84 @@ separate dev team rather than an in-tree Worldtree tool.
 
 ## Current state / in-flight
 
-_As of 2026-05-23 (end of day, post-#12 implementation, pre-commit):_
+_As of 2026-05-23 (end of day, post-v0.2.1 layout fix):_
 
-**Status: issue #12 (presenter contract semantics amendment)
-TDD-complete, in working tree, awaiting commit.** Seven core issues
+**Status: v0.2.0 + v0.2.1 shipped and pushed.** Seven core issues
 complete (`sse_client` #1, `sessions` #2, `cli` #3, `tui` #4,
 `--end-user-id` #5, TUI startup error visibility #6, presenter
 contract semantics amendment #12) + robustness fix #7 (MalformedSseData
-+ empty-skip). 208/208 tests GREEN; ruff clean. pyproject.toml bumped
-to v0.2.0; `uv.lock` refreshed. Working tree has 9 modified files +
-the new `docs/contracts/issues/12.contract.md` (untracked); commit not
-yet authored.
++ empty-skip) + v0.2.1 TUI layout fix on top. 209/209 tests GREEN;
+ruff clean; working tree clean; both v0.2.0 and v0.2.1 tags on
+`origin/main`.
 
 Last commits on `main`:
-- `8282156` snapshot: persistent-memory Heimdall scope-model foot-gun (post-v0.1.0)
-- `804c2df` feat(sessions,cli,tui): issues #5 + #6 + worldtree-dev consumer-API follow-up (tagged v0.1.0)
+- `c85f6bd` fix(tui): anchor layout via dock so Input never moves (v0.2.1)
+- `3b9c610` feat(cli,tui): issue #12 — presenter contract semantics amendment (v0.2.0)
+- `8282156` snapshot: persistent-memory Heimdall scope-model foot-gun
+- `804c2df` feat(sessions,cli,tui): issues #5 + #6 + worldtree-dev follow-up (v0.1.0)
 
-`--send` validated end-to-end against personal Worldtree at v0.1.0
-(`http://10.250.50.152:8081`, mimir on qwen3.6-35-a3b, 2026-05-23 smoke
-returned `[done] turn_id=116 duration_ms=5467`). Lofn smoke is
-**auth-unblocked** as of 2026-05-23 — worldtree-dev confirmed our key
-(`c990f0be`) already covers Tier 1 agents via the `agent.call:*`
-baseline policy; the initial "scope-add needed" diagnosis was a phantom
-ask (see Tried-and-abandoned). The actual lofn fix shipped as issue #5
-(`--end-user-id` flag).
-
-**Async cross-frontier activity in flight:**
-- Issue #12 code-review consult posted to volva 2026-05-23 (althing
-  thread `01KSBH8GYH4G3H03T767X613W7`). Reply pending in inbox.
+**Smoke status:**
+- `--send` v0.1.0 mimir smoke clean earlier today (`[done] turn_id=116
+  duration_ms=5467`).
+- `--send` v0.2.0 mimir smoke survived a Worldtree-side stall (see
+  Tried-and-abandoned for the 2-events-then-silence diagnostic shape).
+  After worldtree-dev's operator-side `:8081` restart, mimir responded
+  cleanly with the new v0.2.0 rendering shape: coalesced thinking
+  runs, `. ` ASCII prefix for demoted telemetry, `duration=11.8s` /
+  `usage 40764 in -> 850 out (41614 total, 0 cached)` formatting.
+- TUI v0.2.0 was visually broken (Input pane bouncing with thinking
+  runs); v0.2.1 fixed via dock-based layout. Operator confirmed
+  "a lot better" interactively.
+- Lofn smoke is **auth-unblocked** (Tier 1 agent.call:* baseline
+  covers it; the phantom-scope-add diagnosis is settled in
+  Tried-and-abandoned). Not yet run.
 
 **Outstanding operator-side todos:**
-- **Commit issue #12 work** + tag v0.2.0 + push. 9 modified files +
-  new `12.contract.md` ready.
-- **Post-v0.2.0 mimir smoke (the visual one)** — `source env.sh && uv
-  run ratatoskr --new --agent mimir --send "test"` to eyeball the new
-  rendering (`. thinking: ...` coalesce, `. worker_phase: ...` demotion,
-  `duration=5.5s` formatting, `usage 6756 in -> 126 out (...)` shape).
-  The v0.1.0 mimir smoke confirmed wire-level backwards compat but
-  did NOT exercise the v0.2.0 rendering.
-- **Post-v0.2.0 lofn smoke** — `source env.sh && uv run ratatoskr
-  --new --agent lofn --send "hello"` (env.sh ships
-  `RATATOSKR_END_USER_ID="ratatoskr-tui"`). Now unblocked on auth.
+- **Post-v0.2.1 lofn smoke** — `source env.sh && uv run ratatoskr
+  --new --agent lofn --send "hello"`. Now unblocked on auth.
+- **Post-v0.2.1 TUI multi-turn eyeball** — confirm thinking-run
+  bouncing is gone across multiple turns; the layout fix has only
+  been confirmed for a single turn so far.
 
 **Pending issues filed but not started:**
 - **Issue #8 (startup agent picker)** — filed but unscaffolded.
-  `GET /agents` is free to call (worldtree-dev confirmed); auth side
-  is unblocked. Depends on #5 composably (both thread through
-  `ParsedArgs` → `_resolve_then_run`).
+  Worldtree-dev confirmed `GET /agents` requires no special scope;
+  auth side is unblocked. Composes with #5 (both thread through
+  `ParsedArgs` → `_resolve_then_run`). **This is the next planned
+  feature** (see Recent decisions §5-sequencing).
 - **Issue #9 (spec-pin refresh v0.19.0 → v0.22.1)** — filed
-  2026-05-23. Documentation debt. None of the v0.20.0/v0.21.0/v0.22.0
-  changes break ratatoskr's existing surface; the pin lies about
-  what we've committed to.
-- **Issue #10 (subject:{type,id} migration)** — filed 2026-05-23 to
-  track Worldtree #196's LOCKED-but-not-shipped breaking change.
-  Don't pre-implement per worldtree-dev's explicit guidance.
+  2026-05-23. Documentation debt; defer unless we need a v0.20.0+
+  capability.
+- **Issue #10 (subject:{type,id} migration)** — filed 2026-05-23
+  to track Worldtree #196. Don't pre-implement per worldtree-dev.
 - **Issue #11 (AdminEvents pane auth prerequisite)** — filed
-  2026-05-23. Future side-pane requires `admin.events.read` scope.
+  2026-05-23. Future side-pane needs `admin.events.read` scope.
 
-Branch: `main` (dirty with #12 work pending commit). Remote:
+**Pending Worldtree-dev follow-up:**
+- worldtree-dev committed (althing `01KSBKTG096Q…`) to file a
+  Worldtree-side issue for the stall-watchdog gap (cancel-check is
+  inside the engine-event loop, so a never-yielding first-LLM-call
+  bypasses the 300s watchdog). Will file after the immediate stall
+  is cleared.
+- Ratatoskr-side companion (potential): a client-side stall watchdog
+  (e.g., 90s-no-events → `[server_stalled]` stderr label, keep
+  connection). Defer until recurrence; defense-in-depth regardless of
+  whether Worldtree fixes its own.
+
+Branch: `main` (clean). Remote:
 `origin → git@gitea.phasefinal.com:vh/ratatoskr.git`.
 
 **Next natural moves:**
 
-1. **Triage volva's #12 code-review** when the reply lands in the
-   inbox; apply tactical fixes inline, surface architectural calls.
-2. **Commit + tag v0.2.0 + push.**
-3. **Post-v0.2.0 smokes** — mimir (visual), lofn (newly unblocked).
-4. **Issue #8 (startup agent picker)** — scaffold + contract, then
-   TDD. Composes with the forward end_user_id direction (see Recent
-   decisions).
-5. **Side-pane issues** — Persona pane first (file-tail, cheap).
-6. **Issue #9 (spec-pin refresh)** — defer unless we need a v0.20.0+
+1. **Lofn smoke** — quick operator-side eyeball.
+2. **Issue #8 (startup agent picker)** — scaffold + contract +
+   TDD. Independent of layout work; gates §5 side-panes work per
+   the 2026-05-23 sequencing decision.
+3. **§5 side-panes work** — Persona pane first per design-brief; the
+   collapsible Thinking pane + Debug pane proposals fold IN as
+   additional `TabbedContent` tabs alongside Persona/Tools/AdminEvents.
+   Reshapes layout from vertical-stack to Horizontal two-column.
+4. **Issue #9 (spec-pin refresh)** — defer unless we need a v0.20.0+
    capability (e.g., `memory_context` for Phase 2.1).
 
 ## Recent decisions
@@ -134,6 +141,8 @@ decision. Captures rationale that won't be obvious from code alone.
 - `[2026-05-23]` **Issue #5 (`--end-user-id`) implemented via TDD.** Small surface change across three modules (sessions, cli, tui): `create_session(client, agent_id, *, end_user_id=None)` widens with optional kwarg; body conditionally adds the field when non-None (INV-002: omitting != sending empty); PRE-003 asserts non-empty. `ParsedArgs.end_user_id: str | None = None` field; `--end-user-id` CLI flag with non-empty validation (mirrors `--send` check). `_amain` and `_resolve_then_run` thread `end_user_id=args.end_user_id` to their `create_session` calls. Post-#6 adjustment: the contract originally named `on_mount` as the TUI threading site, but #6 had moved session resolution to `_resolve_then_run` — same shape, different function. 7 new tests across the 3 modules.
 - `[2026-05-23]` **Worldtree-dev consult landed authoritative consumer-API guidance** (althing thread `01KSBARG2B8M8C82H6AJGJWX1B`). Key takeaways shaped follow-on work: (1) `end_user_id` is a free-form partition key for long-term memory + persona/valence state; same value → same partition, different values → fully isolated. For Vuong-debugging-Worldtree the recommended posture is a project-stable default with `--end-user-id` override. (2) No programmatic `requires_end_user_id` discovery on `GET /agents` — "try and react to 422" remains the pattern. (3) Breaking-change #196 LOCKED but not shipped: `subject:{type,id}` replaces `end_user_id` at future v0.22.x or v0.23.0; don't pre-implement. (4) Spec pin (v0.19.0) is 3 minor versions stale (current v0.22.1); none of v0.20.0/v0.21.0/v0.22.0 break ratatoskr's surface but the pin lies about what we're committed to. (5) User-Agent header: send one (`ratatoskr/<version> (vh@phasefinal.com)`). (6) `agents.call:lofn` scope needed for lofn smoke. (7) `GET /agents` requires no special scope; issue #8 unblocked on auth.
 - `[2026-05-23]` **Follow-up acted on:** User-Agent header added to both `_amain` and `_resolve_then_run` httpx.AsyncClient constructions (with `importlib.metadata` version lookup + fallback to `0.0.0`); `RATATOSKR_END_USER_ID` env-var fallback added to `_parse_args` (resolution: flag > env > None); env.sh ships `RATATOSKR_END_USER_ID="ratatoskr-tui"` as project-stable default. Original issue #5 posture rejected env-var fallback as "papering over isolation"; revised after worldtree-dev's guidance that the realistic single-operator use case wants partition continuity. Issue #5 + #3 contracts amended in-place to document the env-var fallback. Infra-ops pinged via althing for `agents.call:lofn` scope (broker pattern; they forwarded to worldtree-dev). Three Gitea issues filed: #9 (spec-pin refresh), #10 (subject:{type,id} migration tracking), #11 (AdminEvents pane auth prereq).
+- `[2026-05-23]` **v0.2.1 layout fix: dock-anchored TUI chrome so Input never moves** (commit `c85f6bd`, tag `v0.2.1`). Reported during the v0.2.0 mimir TUI smoke: Input bouncing up/down throughout a turn, tokens landing at shifting screen positions. Cause: v0.2.0's `Static(id="thinking-current")` was yielded between `hint` and `Footer` in the auto-stacked vertical flow, so each `display=True/False` toggle per thinking-run shifted Input + identity + hint vertically; RichLog growth from streaming text also drifted Input downward. Fix: `RatatoskrApp.DEFAULT_CSS` docks the chrome to screen edges — `thinking-current` docks top under Header; `transcript` (RichLog) gets `height: 1fr` and absorbs all reflows internally via its scroll viewport; `prompt`, `identity`, `hint` all dock bottom (locked above Footer). Compose order moved `thinking-current` to position 2 (right after Header) so source-order matches the dock layout. **Operator-confirmed "a lot better"** interactively. Pure UI fix; no public API change; tests pass without modification. v0.2.0 → v0.2.1 (patch). I couldn't verify in a TTY from this non-interactive session — the design was sound enough to ship blind, with operator verification post-commit. Going forward: TUI-layout patches like this are "ship + operator verifies" since the TTY is the load-bearing test surface and respx + Pilot mocks can't catch screen-relative positioning bugs.
+- `[2026-05-23]` **Sequencing decision: design-brief §5 side-panes work absorbs the inline collapsible-Thinking-pane + Debug-pane proposals; do issue #8 (startup agent picker) BEFORE §5.** Surfaced during the v0.2.1 follow-up discussion. The operator's proposal — "create a collapsible pane for all thinking tokens; text_boundary goes to a debug pane" — is exactly §5-shaped work (the design-brief proposes a `Horizontal` two-column layout with `TabbedContent` for Persona/Tools/AdminEvents/BifrostState/ServerLog). Building inline-Collapsibles now and then rebuilding as `TabbedContent` panes at §5 would be wasted work. So: do #8 first (independent surface, no layout overlap), then §5 (which folds in Thinking + Debug panes alongside the design-brief's named §5 panes). Interim acceptance: v0.2.1 fixes the structural layout-bouncing pain; transcript-dominated-by-thinking is still real but doesn't degrade further — operator can scroll back, Input doesn't move, tokens land predictably. The interim "noisy transcript" pain is real but bounded; §5 work resolves it cleanly.
 - `[2026-05-23]` **Issue #12 (presenter contract semantics amendment) implemented via TDD.** Headline: thinking deltas render as ONE coalesced growing line (CLI) / one closed RichLog entry per run + live Static(id="thinking-current") widget per-delta (TUI), not 50 lines per turn. Introduced stateful per-turn presenters: `CliPresenterState` (cli.py) and `TuiPresenterState` (tui.py), both `@dataclass(slots=True)` with thinking_buffer + thinking_open (+ text_written_since_newline for CLI). Editorial promotion line settled: load-bearing = Text/Done/Error/Cancelled (no prefix); demoted telemetry = WorkerPhase/Thinking/TextBoundary/ToolStart/ToolResult (CLI `. ` ASCII prefix; TUI `· ` Unicode dim prefix). CLI stdout/stderr newline-boundary INV-005: when text was streamed mid-line, flush a `\n` to stdout before writing terminal labels to stderr; `text_written_since_newline = not event.content.endswith("\n")` per Volva F4 fix. Helpers `_format_duration_ms` (`347ms` / `5.5s` / `1.2m` autoscale) and `_format_usage` (`6756 in -> 126 out (6882 total, 0 cached)` with arrow="->" CLI or "→" TUI). Per Vor (eitri-smithy-dev cross-frontier consult, althing 01KSBE52YZR5) + Volva paraphrase (5 contract-text ambiguities all fixed in #12.contract.md). `[create_session]` lifecycle line demoted to `. create_session:` (written directly by `_amain`, bypasses state.render). Old `_render_event` / `_render_event_to_log` functions and their TestRenderEvent/TestRenderEventToLog classes removed (no-backwards-compat rule). Contracts amended: #3 (CliPresenterState block + `_run_turn` thread state + `_amain` create_session demotion + `_format_*` helper blocks), #4 (TuiPresenterState block + `_stream_turn_worker` state construction + `compose` Static widget addition). 39 new tests; 19 obsolete tests removed; net 208 GREEN. v0.1.0 → v0.2.0 (minor; pre-amendment output shape broken intentionally — scripts grepping `[thinking] '` no longer work; that's the intended cleanup). Cross-frontier design pass with eitri-smithy-dev returned 16-of-16 confirmed decisions + 4 material divergences applied (ASCII `· ` factual fix, RichLog-one-entry-per-run vs inline-mirror, presenter-state object vs stateless, "contract semantics amendment" framing not "polish"). Calibration note: eitri-smithy-dev's value here was *architectural* (state-object pattern + chronological-vs-live decoupling) not just *tactical*; the framing rename alone justified the consult. Volva paraphrase round added 5 prose-precision fixes (INV-001 "growing display" semantics, TUI hide mechanism unification, render_error security/readability tension, newline-tracking corner case, [create_session] integration path).
 - `[2026-05-23]` **Forward direction: Ratatoskr will require `end_user_id` for EVERY access before too long.** Operator's call. Reasoning: even Tier 1 foundational agents (mimir, all Asgardians) that don't *require* `end_user_id` server-side currently fall back to a `_no_end_user` sentinel substrate partition — effectively pollution from a single-operator-debug-tool's perspective. The right shape is "every conversation has an explicit partition key." `RATATOSKR_END_USER_ID="ratatoskr-tui"` env-default in env.sh is the first step toward that posture; once we've validated the partition-isolation experience, the next move is making `end_user_id` mandatory (probably remove the `None`-default in `_parse_args`, fail-closed with a UsageError if neither flag nor env provides it). Consequence for cross-project asks: declined worldtree-dev's offer to ship `requires_end_user_id: bool` on `AgentInfoResponse` because we'd treat every value as true regardless; the try-and-react-to-422 pattern goes away from our side because we never send a request without the field. File a ratatoskr issue when scheduling the change — touches `_parse_args` validation + `_resolve_then_run` + `_amain` + tests + contract amendments to #3 / #5. Treat as a v0.2.0 minor (breaking: existing `--new --agent mimir` without env or flag would start failing). **Cross-frontier alignment (worldtree-dev ack 2026-05-23, althing 01KSBD9FPMCWJMBXNNS4B3MYBS):** the platform side agrees with this framing — `_no_end_user` is a substrate accommodation for identity-less transports, NOT a consumer model. The fallback's `_is_fallback=True` trap door (#185 INV-185-5/8) "could become operator-controlled later" per worldtree-dev, meaning Worldtree itself may tighten the substrate-fallback path. Ratatoskr's forward posture pre-empts that tightening — moving from "we send end_user_id when set" to "we never send a request without end_user_id" stays consumer-correct regardless of what Worldtree does with the fallback knob.
 
@@ -153,4 +162,5 @@ defense against re-attempting the same cul-de-sac.
 - `[2026-05-21]` **TUI session-identity rendering via `self.sub_title` + `self.hint` plain attributes.** Stored state but never rendered to a visible widget. The contract's "session-identity-always-visible" invariant was satisfied at the state-attribute level but not the user-visible-widget level. Tests asserted the attributes (which passed); Volva code-review flagged the gap. Fix: dedicated `Static(id="identity")` + `Static(id="hint")` widgets in compose; `_set_hint()` helper mirrors state → widget. Calibration evidence for the "TDD catches state, code-review catches whether the user can see it" pattern.
 - `[2026-05-23]` **Using the cross-model review agent's name directly in composed prose.** The peer review agent's name (the `althing` handle starting with "V-o-l-v-a") is one letter from a body-part term. Anthropic's content classifier does fuzzy matching and intermittently blocks responses mid-stream when the name appears in composed prose sentences (especially in meta-commentary about the agent's work). Direct-quoted tool output (e.g., the `althing-cli thread` body) passes through fine. Mitigation: use role descriptions ("the cross-model reviewer," "the paraphrase peer") in prose rather than the name; quote content via tool output. Confirmed by switching to Sonnet 4.6 for a test read — same raw content read cleanly when fetched via Bash rather than composed into an LLM response. This is a persistent environmental constraint, not a one-off.
 - `[2026-05-22]` **`json.loads(sse.data)` unguarded against empty data.** `_iter_events` unconditionally called `json.loads` on every dispatched `ServerSentEvent`. When `httpx_sse` surfaced a frame with `id:` present but `data:` empty (a known library-vs-spec divergence — RFC says don't dispatch; httpx_sse is permissive), `json.loads('')` raised `JSONDecodeError` → propagated through Textual's worker → app crash. Crashed mimir conversation at turn 93/seq 1078 after 1077 successful events. Fix: `if sse.data == '': continue` BEFORE `_parse_sse_id` (empty-data event with a malformed id is still a keepalive — don't reorder). Non-empty malformed data raises new `MalformedSseData(raw[:200])`. Don't reintroduce unconditional `json.loads(sse.data)`; always pre-check for the empty case.
+- `[2026-05-23]` **Diagnostic shorthand: "2-events-then-silence" = Worldtree-side LLM-call wedge, not ratatoskr.** If a mimir `--send` smoke shows exactly two stderr events — `. create_session: ...` followed by `. worker_phase: phase=BuildingPrompt ...` — and then nothing for >60s, the root cause is upstream of ratatoskr. Worldtree's `service.py:2560` gates the `CallingLLM` event on the engine yielding its first LLM-provider chunk; if that provider connection is wedged at the TCP level, the `async for` never iterates and the SSE stream stays silent forever. ratatoskr's `read=None` httpx timeout (the issue #1 + #4 INV-007 fix for "5s default killed mid-stream during mimir's thinking") waits patiently as designed; there's no client-side stall watchdog above the read-timeout layer. Worldtree's OWN stall watchdog (300s `_start_stall_timer`) exists but its cancel-check is INSIDE the engine-event loop, so a never-yielding first-LLM-call bypasses it. Confirmed by worldtree-dev (althing thread `01KSBKTG096Q07JVRG41JXA1DD`). **Don't waste time bisecting ratatoskr code when this shape appears** — diagnose the LLM-provider connection state at Worldtree's host. Restarting the Worldtree service (`:8081` in our case) cleared a wedged llama-swap connection. Future ratatoskr issue worth filing if recurrence: client-side stall watchdog (e.g., 90s-no-events → `[server_stalled]` stderr label, keep connection open). Also worth knowing: 10.250.50.152 hosts 3 Worldtree instances (`:8080`, `:8081`, `:8082`) — each with its own DB and key namespace. Our key is valid only on `:8081`.
 - `[2026-05-23]` **Phantom "per-Tier-1-agent scope add" pattern.** Issue #5's lofn 422 was initially diagnosed (with worldtree-dev's first reply) as needing `agents.call:lofn` added to ratatoskr's existing key. Routed through infra-ops via althing per the credential-brokerage rule; infra-ops discovered no public scope-mutation endpoint on personal Worldtree, brokered to worldtree-dev for the actual mechanism. Worldtree-dev came back with a correction: their first answer conflated two distinct Heimdall scope namespaces. **Tier 1 foundational agents** (mimir, lofn, soong, all Asgardians) are covered by a blanket `agent.call:*` (singular) baseline rule in `config/policies.yaml > tiers.<tier>.scopes` for ALL authenticated tiers including `user`. There is no per-agent grant for Tier 1 — the baseline rule covers it. **Tier 3 consumer-defined agents** (IDs containing `:`, like `vh:custom-bot`) use the plural `agents.call:<owner>:<agent>` shape granted implicitly via owning a `consumer_agents` DB row, registered through `POST /agents/define`. The two notations differ by one letter and that was the source of the confusion. **The actual lofn fix was issue #5's `--end-user-id` flag — it was always a request-body validation, not an auth-scope gate.** Don't ping infra-ops for "per-Tier-1-agent scope adds" again; the pattern is a phantom ask. Real future infra-ops asks: admin-tier key for the AdminEvents pane (`admin.events.read` scope, different tier), and Tier 3 custom-agent registration (different flow entirely, requires `POST /agents/define`).