A 2026-05-23 mimir smoke against personal Worldtree (post-issues #5 + #6 commit 804c2df / v0.1.0) exposed multiple presenter UX issues that materially degrade the observability surface ratatoskr is supposed to provide.
Headline problem: thinking tokens render one-per-line. Worldtree streams thinking as token-deltas (just like text), but the current renderers (cli._render_event and tui._render_event_to_log) write [thinking] '<token>'\n per delta. A 50-token thinking phase = 50 lines of [thinking] spam.
Broader scope: every demoted-telemetry event renders with the same visual weight as actual model output, drowning the load-bearing signal. Specific issues:
[done] on stderr interleaves with stdout text in a real terminal
Mission constraint
Per persistent-memory: "the product IS the observability surface; chat is the input mechanism." Suppressing events is the wrong direction. The right direction is rendering them cleanly, with editorial judgment about what's load-bearing transcript (Text / Done / Error / Cancelled) vs demoted telemetry (WorkerPhase / Thinking / TextBoundary / ToolStart / ToolResult).
A refined mission lens emerged from this smoke: "the observability surface presented at the right zoom level for the operator's task." Per-event raw streaming is the wrong zoom for normal debugging; coalesced thinking + demoted telemetry is the right zoom.
Scope
Both presenters (cli --send + Textual TUI).
Editorial demotion of telemetry events.
Thinking-coalesce (single growing line in cli; single growing widget in tui).
Duration + usage formatting.
TTY-interleave fix between stdout (text) and stderr (meta).
Dedicated TUI Static(id="thinking-current") widget for live thinking state (chronological-in-RichLog + always-visible-in-widget = two views same data).
Out of scope
New public flags (--quiet, --verbose, --debug). Decision: ship flag-less; revisit if a real pipe-to-jq use case appears.
Side-pane work (AdminEvents, Persona, Tools, BifrostState, ServerLog per design-brief §5). The dedicated thinking widget added here is the foundation, but it's not a "Persona pane" or "AdminEvents pane" yet.
Cross-process resume / transcript persistence. Per design-brief §8d, deferred to v2.
Design pass
Sent the design questions to eitri-smithy-dev for a cross-frontier-model technical pass before contract authoring. Thread 01KSBE52YZR5E3SPTKA672JE43. Tactical calls (closure trigger, formatting, edge cases) will be applied directly to the contract; architectural-shape calls (mission refinement, widget yes/no, flag policy, editorial promotion line) will route through the operator.
Acceptance
Two parallel contract amendments to issues #3 (cli) and #4 (tui) landing in one commit.
Mid-conversation mimir smoke produces a transcript where:
Thinking renders as one coalesced growing line/widget (not N lines).
Telemetry events visually demoted (· prefix in cli, dim style in tui).
duration_ms displayed as 5.5s / 347ms / 1.2m.
usage displayed as natural-language flow.
[done] line lands on its own line after a clean stdout flush.
TUI: dedicated Static(id="thinking-current") widget shows current thinking state, mirrors content into RichLog as chronological dimmed entries.
All existing tests + new rendering tests GREEN under uv run pytest tests/.
uv run ruff check src/ tests/ clean.
Version posture
v0.1.0 → v0.2.0 (minor). Output shape changes; scripts grepping [thinking] '<token>' would break. Pre-v1.x means breaking changes are minor.
## Context
A 2026-05-23 mimir smoke against personal Worldtree (post-issues #5 + #6 commit `804c2df` / v0.1.0) exposed multiple presenter UX issues that materially degrade the observability surface ratatoskr is supposed to provide.
**Headline problem:** thinking tokens render one-per-line. Worldtree streams thinking as token-deltas (just like text), but the current renderers (`cli._render_event` and `tui._render_event_to_log`) write `[thinking] '<token>'\n` per delta. A 50-token thinking phase = 50 lines of `[thinking]` spam.
**Broader scope:** every demoted-telemetry event renders with the same visual weight as actual model output, drowning the load-bearing signal. Specific issues:
- thinking deltas: 50-line spam (above)
- text_boundary: appears inline mid-text, breaking sentences visually
- worker_phase: in-transcript with no visual demotion
- ToolStart / ToolResult: same telemetry-vs-transcript tension
- duration_ms / usage dict: unformatted (`duration_ms=5467`, `usage={'prompt_tokens': ..., ...}`)
- [done] on stderr interleaves with stdout text in a real terminal
## Mission constraint
Per persistent-memory: "the product IS the observability surface; chat is the input mechanism." Suppressing events is the wrong direction. The right direction is rendering them cleanly, with editorial judgment about what's load-bearing transcript (Text / Done / Error / Cancelled) vs demoted telemetry (WorkerPhase / Thinking / TextBoundary / ToolStart / ToolResult).
A refined mission lens emerged from this smoke: "the observability surface presented at the right zoom level for the operator's task." Per-event raw streaming is the wrong zoom for normal debugging; coalesced thinking + demoted telemetry is the right zoom.
## Scope
- Both presenters (cli `--send` + Textual TUI).
- Editorial demotion of telemetry events.
- Thinking-coalesce (single growing line in cli; single growing widget in tui).
- Duration + usage formatting.
- TTY-interleave fix between stdout (text) and stderr (meta).
- Dedicated TUI `Static(id="thinking-current")` widget for live thinking state (chronological-in-RichLog + always-visible-in-widget = two views same data).
## Out of scope
- New public flags (`--quiet`, `--verbose`, `--debug`). Decision: ship flag-less; revisit if a real pipe-to-jq use case appears.
- Side-pane work (AdminEvents, Persona, Tools, BifrostState, ServerLog per design-brief §5). The dedicated thinking widget added here is the foundation, but it's not a "Persona pane" or "AdminEvents pane" yet.
- i18n / locale-aware duration formatting. English-only Vuong-only.
- Cross-process resume / transcript persistence. Per design-brief §8d, deferred to v2.
## Design pass
Sent the design questions to eitri-smithy-dev for a cross-frontier-model technical pass before contract authoring. Thread `01KSBE52YZR5E3SPTKA672JE43`. Tactical calls (closure trigger, formatting, edge cases) will be applied directly to the contract; architectural-shape calls (mission refinement, widget yes/no, flag policy, editorial promotion line) will route through the operator.
## Acceptance
- Two parallel contract amendments to issues #3 (cli) and #4 (tui) landing in one commit.
- Mid-conversation mimir smoke produces a transcript where:
- Thinking renders as one coalesced growing line/widget (not N lines).
- Telemetry events visually demoted (`· ` prefix in cli, dim style in tui).
- duration_ms displayed as `5.5s` / `347ms` / `1.2m`.
- usage displayed as natural-language flow.
- [done] line lands on its own line after a clean stdout flush.
- TUI: dedicated `Static(id="thinking-current")` widget shows current thinking state, mirrors content into RichLog as chronological dimmed entries.
- All existing tests + new rendering tests GREEN under `uv run pytest tests/`.
- `uv run ruff check src/ tests/` clean.
## Version posture
v0.1.0 → v0.2.0 (minor). Output shape changes; scripts grepping `[thinking] '<token>'` would break. Pre-v1.x means breaking changes are minor.
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Context
A 2026-05-23 mimir smoke against personal Worldtree (post-issues #5 + #6 commit
804c2df/ v0.1.0) exposed multiple presenter UX issues that materially degrade the observability surface ratatoskr is supposed to provide.Headline problem: thinking tokens render one-per-line. Worldtree streams thinking as token-deltas (just like text), but the current renderers (
cli._render_eventandtui._render_event_to_log) write[thinking] '<token>'\nper delta. A 50-token thinking phase = 50 lines of[thinking]spam.Broader scope: every demoted-telemetry event renders with the same visual weight as actual model output, drowning the load-bearing signal. Specific issues:
duration_ms=5467,usage={'prompt_tokens': ..., ...})Mission constraint
Per persistent-memory: "the product IS the observability surface; chat is the input mechanism." Suppressing events is the wrong direction. The right direction is rendering them cleanly, with editorial judgment about what's load-bearing transcript (Text / Done / Error / Cancelled) vs demoted telemetry (WorkerPhase / Thinking / TextBoundary / ToolStart / ToolResult).
A refined mission lens emerged from this smoke: "the observability surface presented at the right zoom level for the operator's task." Per-event raw streaming is the wrong zoom for normal debugging; coalesced thinking + demoted telemetry is the right zoom.
Scope
--send+ Textual TUI).Static(id="thinking-current")widget for live thinking state (chronological-in-RichLog + always-visible-in-widget = two views same data).Out of scope
--quiet,--verbose,--debug). Decision: ship flag-less; revisit if a real pipe-to-jq use case appears.Design pass
Sent the design questions to eitri-smithy-dev for a cross-frontier-model technical pass before contract authoring. Thread
01KSBE52YZR5E3SPTKA672JE43. Tactical calls (closure trigger, formatting, edge cases) will be applied directly to the contract; architectural-shape calls (mission refinement, widget yes/no, flag policy, editorial promotion line) will route through the operator.Acceptance
·prefix in cli, dim style in tui).5.5s/347ms/1.2m.Static(id="thinking-current")widget shows current thinking state, mirrors content into RichLog as chronological dimmed entries.uv run pytest tests/.uv run ruff check src/ tests/clean.Version posture
v0.1.0 → v0.2.0 (minor). Output shape changes; scripts grepping
[thinking] '<token>'would break. Pre-v1.x means breaking changes are minor.Presenter rendering polish: thinking-coalesce + telemetry demotion + formattingto Presenter contract semantics amendment: stateful event coalescing + telemetry demotionClosed — shipped. Presenter contract semantics amendment (stateful coalescing + telemetry demotion) implemented via TDD (commit
3b9c610, v0.2.0). CliPresenterState + TuiPresenterState; thinking deltas coalesce to one growing line.