Implement list_runs in terms of list_runs_with_projection and drop the
now-unused RunDatabase::build_summary wrapper. Removes the duplicated
catalog-iteration loop and sort key.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Thread the RunProjection already built by SlateDB list_runs through to the board runs handler, so PR, sandbox, and pending-question metadata are read without reopening each run.
## Summary
Run cancellation now reaches in-flight agent work instead of waiting for
an agent stage to finish or recording cancellation as a failed stage.
The workflow cancellation primitive is now
`tokio_util::sync::CancellationToken`, with child tokens passed through
setup, handlers, manager-loop child runs, sandbox streaming commands,
CLI agent invocations, and API agent sessions.
### Plan Summary
- Promote run cancellation to `CancellationToken` while keeping stall
timeout separate.
- Route CLI agents through cancellable sandbox streaming with optional
timeouts.
- Bridge run cancellation into API sessions and preserve
`Error::Cancelled` propagation.
- Add typed events/projections for CLI cancellation and timeout.
## Cancellation flow
```mermaid
flowchart TB
RunToken[Run CancellationToken]
Executor[Core executor]
Services[RunServices]
Manager[Manager-loop child run]
CLI[Agent CLI backend]
API[Agent API backend]
Sandbox[Sandbox streaming exec]
Session[fabro-agent Session]
RunToken --> Executor
RunToken --> Services
Services -- child_token --> Manager
Services -- child_token --> CLI
CLI -- child_token --> Sandbox
Services --> API
API -- bridge guard --> Session
```
## What changed and why
- `RunOptions`, `RunServices`, core `ExecutorOptions`, CLI/server run
state, and detached-run guards now use `CancellationToken` instead of
`Arc<AtomicBool>`. Dropping services or tokens still does not mean
cancellation; only explicit `.cancel()` does.
- Manager-loop child workflows are given child tokens so parent
cancellation propagates down, while stop/max-cycle cancellation remains
scoped to the child workflow.
- Stall timeout remains intentionally separate as a stall token and
still returns `Error::StallTimeout { node_id }`, not `Error::Cancelled`.
- Agent, prompt, human, fan-in, and parallel handler paths now pass
cancellation tokens through and avoid converting `Error::Cancelled` into
normal failed outcomes.
## Agent backend behavior
CLI-mode agents no longer launch detached `setsid` jobs with temp
stdout/stderr/exit-code polling. They run through
`Sandbox::exec_command_streaming` with a child token; a missing node
timeout passes `None` to preserve the existing unbounded agent runtime,
while explicit node timeouts still apply. Cancelled CLI runs emit
`agent.cli.cancelled`, clean temp files, and return `Error::Cancelled`;
timed-out CLI runs emit `agent.cli.timed_out` and return a handler
timeout error; `agent.cli.completed` remains natural-exit only.
API-mode agents install a per-invocation `SessionCancelBridgeGuard`
after acquiring a fresh or cached session. The guard maps the run token
into the session interrupt reason and session cancel token, and aborts
stale bridge tasks before session replacement or cache reinsertion so
reused sessions are not tied to old run tokens. `Session::initialize`
now returns `Result`, and project-doc, skill, MCP, and environment
discovery paths check cancellation and pass child tokens to sandbox
commands.
## Sandbox and event model
`Sandbox::exec_command_streaming` now accepts `Option<u64>` for timeout.
Production streaming implementations use a pending future for `None`
instead of a giant sleep, while the trait fallback maps `None` to
`u64::MAX` only when delegating to non-streaming `exec_command`.
The run event model now includes typed `agent.cli.cancelled` and
`agent.cli.timed_out` payloads with stdout, stderr, and duration, plus
conversion and projection support. OpenAPI/client regeneration was
unnecessary because the API schema already models run events with a free
event string and arbitrary properties; only Rust event types changed.
## Reviewer notes
Expect signature churn around `Session::initialize`,
`CodergenBackend::run`, `RunOptions.cancel_token`,
`StartServices.cancel_token`, and `Sandbox::exec_command_streaming`. The
main behavioral checks are that user cancellation reaches in-flight
CLI/API work and that timeout/stall paths remain distinct from user
cancellation.
### Fabro Details
<details>
<summary>Ran 9 stages in 117m 40s for $150.32</summary>
| Stage | Duration | Cost | Retries |
|---|---|---|---|
| start | 0s | – | 0 |
| toolchain | 1s | – | 0 |
| preflight_compile | 2m 8s | – | 0 |
| preflight_lint | 2m 13s | – | 0 |
| implement | 77m 12s | $56.78 | 0 |
| simplify_opus | 18m 5s | $5.83 | 0 |
| simplify_gpt | 15m 33s | $87.71 | 0 |
| verify | 1m 48s | – | 0 |
| fmt | 2s | – | 0 |
| **Total** | **117m 40s** | **$150.32** | **0** |
</details>
<details>
<summary>Ran <code>ImplementPlan.fabro</code> (12 nodes and 15
edges)</summary>
```dot
digraph ImplementPlan {
graph [
goal="Implement and simplify",
model_stylesheet="
* { model: claude-opus-4-7; }
"
]
rankdir=LR
start [shape=Mdiamond, label="Start"]
exit [shape=Msquare, label="Exit"]
toolchain [label="Toolchain", shape=parallelogram, script="command -v cargo >/dev/null || { curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y && sudo ln -sf $HOME/.cargo/bin/* /usr/local/bin/; }; cargo --version 2>&1", max_retries=0]
preflight_compile [label="Preflight Compile", shape=parallelogram, script="cargo check -q --workspace 2>&1", max_retries=0]
preflight_lint [label="Preflight Lint", shape=parallelogram, script="cargo +nightly-2026-04-14 clippy -q --workspace --all-targets -- -D warnings 2>&1", max_retries=0]
fix_lints [label="Fix Lints", prompt="The preflight lint step failed. Read the build output from context and fix all clippy lint warnings.", max_visits=3]
implement [label="Implement", prompt="Read the plan file referenced in the goal and implement every step. Make all the code changes described in the plan. Use red/green TDD."]
simplify_opus [label="Simplify (Opus)", prompt="@prompts/simplify.md"]
simplify_gpt [label="Simplify (GPT-55)", prompt="@prompts/simplify.md", model="gpt-55"]
verify [label="Verify", shape=parallelogram, script="cargo +nightly-2026-04-14 clippy -q --workspace --all-targets -- -D warnings 2>&1 && cargo nextest run --cargo-quiet --workspace --status-level fail 2>&1 && cargo dev docs refresh 2>&1 && cargo dev docs check 2>&1", goal_gate=true, retry_target="fixup"]
fixup [label="Fixup", prompt="The verify step failed. Read the build output from context and fix all clippy lint warnings, test failures, and generated docs errors.", max_visits=3]
fmt [label="Format", shape=parallelogram, script="cargo +nightly-2026-04-14 fmt --all 2>&1", max_retries=0]
start -> toolchain
toolchain -> preflight_compile [condition="outcome=succeeded"]
toolchain -> exit
preflight_compile -> preflight_lint [condition="outcome=succeeded"]
preflight_compile -> exit
preflight_lint -> implement [condition="outcome=succeeded"]
preflight_lint -> fix_lints
fix_lints -> preflight_lint
implement -> simplify_opus -> simplify_gpt -> verify
verify -> fmt [condition="outcome=succeeded"]
verify -> fixup
fixup -> verify
fmt -> exit
}
```
</details>
⚒️ Generated with [Fabro](https://fabro.sh)
---------
Co-authored-by: Fabro <noreply@fabro.sh>
Co-authored-by: Bryan Helmkamp <bryan@brynary.com>
Use to_rfc3339_opts with SecondsFormat::Secs so auth status output
shows clean second-precision timestamps instead of nanoseconds.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
### Summary
Billing and stage lists now use the event-sourced `RunProjection` as
their source of truth, so running and retrying stages appear immediately
and runtimes keep advancing in the UI. This removes the checkpoint
completed-node bypass that hid in-flight work and froze totals until the
next server response.
### Plan Summary
- Store stage `started_at`, terminal `duration_ms`, server-internal
`usage`, and lifecycle `state` on `StageProjection`.
- Populate those fields from stage lifecycle events, including retry
transitions and per-attempt reset on new starts.
- Render `/runs/{id}/stages` and `/runs/{id}/billing` from
`RunProjection.iter_stages()`.
- Expose the new API/client fields and tick in-flight billing runtimes
on the web UI.
```mermaid
flowchart TB
Events["Stage lifecycle events"] --> Projection["RunProjection StageProjection"]
Projection --> StagesAPI["GET /runs/{id}/stages"]
Projection --> BillingAPI["GET /runs/{id}/billing"]
StagesAPI --> StageUI["Stage sidebar/stages view"]
BillingAPI --> BillingUI["Billing tab live totals"]
```
### Key decisions
Retry and revisit handling stays one row per node id: latest visit data
wins, while first-seen event sequence keeps ordering stable with
finalize output. `state` is stored rather than derived so `Retrying` is
representable, and old serialized projections still work through the
`effective_state()` fallback. Billing `usage` remains server-internal
and is skipped on the wire; public schemas only expose the fields needed
by `/stages`, `/billing`, and the frontend live timer.
Added focused reducer, server retry/revisit, API round-trip, billing UI,
and event invalidation coverage.
⚒️ Generated with [Fabro](https://fabro.sh)
---------
Co-authored-by: Fabro <noreply@fabro.sh>
Co-authored-by: Bryan Helmkamp <bryan@brynary.com>
### Summary
Silent fallback paths now emit stable warnings instead of degrading
without a user-visible signal. The fallback behavior is unchanged; runs
still continue, but worktree, Git, checkpoint, and LLM failover issues
now show up in the run feed and logs.
### Plan Summary
- Emit run notices for workflow Git/worktree fallback paths.
- Reuse the existing failover event for one-shot LLM provider fallback.
- Add tracing for sandbox pipe drain failures.
### What changed
- Added `worktree_skipped_no_git` and gated `sandbox_git_unavailable`
notices during initialization.
- Added `git_push_failed` and `parallel_base_checkpoint_failed` notices,
including redacted output tails where available.
- Logged GitHub token mint failures with a structured `error` field
before the existing notice.
- Plumbed `Emitter` and `StageScope` through `CodergenBackend::one_shot`
so the API backend emits the existing `agent.failover` event instead of
a duplicate tracing-only warning.
- Extracted sandbox pipe draining into a helper that warns on
stdout/stderr read failures, with unit coverage for the error path.
- Updated CLI snapshots for the new worktree warning in stderr and JSON
event output.
⚒️ Generated with [Fabro](https://fabro.sh)
---------
Co-authored-by: Fabro <noreply@fabro.sh>
Co-authored-by: Bryan Helmkamp <bryan@brynary.com>
## Summary
Stage detail now loads activity from a canonical stage-scoped events
endpoint instead of falling back to the first 1000 run-wide events. This
fixes empty panes for late stages in long runs and removes the
presentation-shaped `StageTurn` API from the wire.
### Plan Summary
- Add `GET /runs/{id}/stages/{stageId}/events` with cursor pagination
and server-side `node_id` filtering.
- Replace frontend stage-turn/fallback loading with paginated
stage-events loading and local event-to-activity projection.
- Broaden SSE/SWR invalidation so every activity event consumed by the
reducer refreshes the per-stage cache.
- Remove `StageTurn` schemas/client models and update demo fixtures plus
pagination/handler/reducer tests.
## What changed and why
The store now scans the run event prefix and filters by `node_id` before
applying the `limit + 1` cutoff. That preserves sparse late-stage
matches that would otherwise be dropped if we reused the run-wide
limited scan and filtered afterward. The real-mode handler returns an
empty page for an unknown stage id in an existing run, while preserving
404 for missing runs.
On the frontend, `run-stages` fetches all pages for the selected stage
and feeds them through `eventsToActivity`, keeping `TurnType` as a local
presentation model. Invalidation now targets `runs.stageEvents(runId,
stageId)` for lifecycle and reducer-consumed activity events
(`stage.prompt`, agent messages/tools, and command events), so active
panes refresh from the existing run event subscription.
The OpenAPI document and generated TS client now expose
`listStageEvents` and drop stale `StageTurn` models. Demo mode serves a
`detect-drift` stage-events fixture using the same cursor semantics as
the real endpoint.
## API notes
`/runs/{id}/stages/{stageId}/turns` is removed; clients should use
`/runs/{id}/stages/{stageId}/events?since_seq=&limit=` and project
events locally. The `stageId` path segment for this endpoint is the
workflow node id, not the visit-qualified `node_id@visit` form used by
command logs/artifacts.
⚒️ Generated with [Fabro](https://fabro.sh)
---------
Co-authored-by: Fabro <noreply@fabro.sh>
Co-authored-by: Bryan Helmkamp <bryan@brynary.com>
Return named PR content from the builder and keep title/body fallback logic inside the builder.
Move the PR body prompt to markdown and scale prompt truncation from model context windows.
Keep PR creation resilient when generated bodies are empty by emitting a reviewer-visible skeleton body.
Make BoardColumnDefinition.id reference the existing BoardColumn schema and carry that typed contract through generated TypeScript, server responses, demo data, and the runs board UI.
Reset coordinator state when the last subscriber leaves, clear pending debounce timers on close, and keep coordinated EventSource construction owned by the coordinator while fallback subscriptions keep their local factories.
Use unknown.ts helpers in parseMessage, factor out parseLeaderPair/Triple
and per-variant parsers to remove repeated typeof guards. Extract
leaderIsFresh() for the staleness check used in three places, and make
RecentEventCache amortized O(1) by walking expired entries from the
oldest instead of scanning the whole map per event. Drop the
closeOnTerminal parameter in run-events; the fallback path computes
close at its single call site.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Prune stale election candidates as generations advance, reset coordination availability on explicit close, and keep fallback subscribers tracked so coordinator shutdown can clean them up consistently.
Stop coordinated election and leadership work when BroadcastChannel posting fails, so tabs degrade cleanly to per-subscriber fallback without stale resync or heartbeat side effects. Expand election coverage for the edge cases called out in the coordination plan.
Elect a single browser tab to own the global attach stream and broadcast run events to sibling tabs. Keep the existing per-tab EventSource path as the fallback when cross-tab coordination is unavailable.
Submitted and Queued lifecycle statuses now live in a dedicated Queued
column rendered to the left of Initializing; Starting stays in
Initializing. The column is omitted from the board when it has no items
so day-to-day boards stay compact.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Render kanban column shells while board/auth/system queries load, so the
"Your runs will appear here" panel no longer flashes before runs arrive.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This change makes bulk `fabro model test` run configured model checks
concurrently instead of serially. A new `--jobs/-j` flag (defaulting to
4, minimum 1) controls the concurrency bound; the single-model path
(`--model <MODEL>`) is unaffected. Under the hood, the serial `for` loop
over configured models is replaced with a
`futures::stream::buffer_unordered(jobs)` pipeline that clones the
shared-state `Client` per request. Completed results carry their
original list index and are sorted before rendering, so final stdout
table rows and JSON output remain in listing order regardless of which
requests finish first.
Three new integration tests verify the concurrency behavior using an
inline Axum harness with a `ConcurrencyGate` barrier. The gate holds all
in-flight requests until the expected number arrive simultaneously, then
releases them, letting tests assert `max_in_flight` exactly rather than
relying on timing. The ordering test goes further by assigning
reverse-listing response delays so the last-listed model always finishes
first; if the index sort were dropped, the JSON result order would
invert and the assertion would fail. A 15-second gate timeout ensures a
regression to serial execution surfaces as a clear `max_in_flight == 1`
failure rather than a hung test.
Existing behavior is fully preserved: unconfigured models are still
skipped without a POST, a configured model returning `skip` after
listing is still a failure, `--deep` uses the same `--jobs` value, and
`--jobs 1` reproduces the previous serial behavior for users hitting
provider rate limits.
### Fabro Details
<details>
<summary>Ran 9 stages in 30m 52s for $19.61</summary>
| Stage | Duration | Cost | Retries |
|---|---|---|---|
| start | 0s | – | 0 |
| toolchain | 1s | – | 0 |
| preflight_compile | 2m 8s | – | 0 |
| preflight_lint | 2m 21s | – | 0 |
| implement | 8m 56s | $3.87 | 0 |
| simplify_opus | 7m 55s | $1.43 | 0 |
| simplify_gpt | 6m 58s | $14.32 | 0 |
| verify | 1m 49s | – | 0 |
| fmt | 2s | – | 0 |
| **Total** | **30m 52s** | **$19.61** | **0** |
</details>
<details>
<summary>Ran <code>ImplementPlan.fabro</code> (12 nodes and 15
edges)</summary>
```dot
digraph ImplementPlan {
graph [
goal="Implement and simplify",
model_stylesheet="
* { model: claude-opus-4-7; }
"
]
rankdir=LR
start [shape=Mdiamond, label="Start"]
exit [shape=Msquare, label="Exit"]
toolchain [label="Toolchain", shape=parallelogram, script="command -v cargo >/dev/null || { curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y && sudo ln -sf $HOME/.cargo/bin/* /usr/local/bin/; }; cargo --version 2>&1", max_retries=0]
preflight_compile [label="Preflight Compile", shape=parallelogram, script="cargo check -q --workspace 2>&1", max_retries=0]
preflight_lint [label="Preflight Lint", shape=parallelogram, script="cargo +nightly-2026-04-14 clippy -q --workspace --all-targets -- -D warnings 2>&1", max_retries=0]
fix_lints [label="Fix Lints", prompt="The preflight lint step failed. Read the build output from context and fix all clippy lint warnings.", max_visits=3]
implement [label="Implement", prompt="Read the plan file referenced in the goal and implement every step. Make all the code changes described in the plan. Use red/green TDD."]
simplify_opus [label="Simplify (Opus)", prompt="@prompts/simplify.md"]
simplify_gpt [label="Simplify (GPT-55)", prompt="@prompts/simplify.md", model="gpt-55"]
verify [label="Verify", shape=parallelogram, script="cargo +nightly-2026-04-14 clippy -q --workspace --all-targets -- -D warnings 2>&1 && cargo nextest run --cargo-quiet --workspace --status-level fail 2>&1 && cargo dev docs refresh 2>&1 && cargo dev docs check 2>&1", goal_gate=true, retry_target="fixup"]
fixup [label="Fixup", prompt="The verify step failed. Read the build output from context and fix all clippy lint warnings, test failures, and generated docs errors.", max_visits=3]
fmt [label="Format", shape=parallelogram, script="cargo +nightly-2026-04-14 fmt --all 2>&1", max_retries=0]
start -> toolchain
toolchain -> preflight_compile [condition="outcome=succeeded"]
toolchain -> exit
preflight_compile -> preflight_lint [condition="outcome=succeeded"]
preflight_compile -> exit
preflight_lint -> implement [condition="outcome=succeeded"]
preflight_lint -> fix_lints
fix_lints -> preflight_lint
implement -> simplify_opus -> simplify_gpt -> verify
verify -> fmt [condition="outcome=succeeded"]
verify -> fixup
fixup -> verify
fmt -> exit
}
```
</details>
⚒️ Generated with [Fabro](https://fabro.sh)
---------
Co-authored-by: Fabro <noreply@fabro.sh>
Two follow-ups that were still costing ~1s per credential probe:
- Bumped the daytona-sdk-rust pin to fa4870f, which deletes a dead
underscore-prefixed _http_client field on Client. The field was
unused but new_with_config built a fresh reqwest::Client for it on
every call, paying the macOS proxy-discovery tax even with our
injection seam in place.
- build_api_keys_configuration was using Configuration::new() and then
overwriting cfg.client with our injected client. The Default impl
generated by openapi-generator builds a reqwest::Client::new() for
the client field eagerly, which we then threw away — another
~470ms hit per probe. Construct the Configuration as a struct
literal so the injected client is the only one we ever build.
Drops the three credential-probe tests from ~700ms to ~10ms.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Routes the two reqwest clients in the Daytona credential probe through
fabro_http (system-proxy) in production and fabro_test::test_http_client
(no_proxy) in tests, by threading an http_client parameter through
check_daytona_api_key_with and build_api_keys_configuration. Bumps the
daytona-sdk-rust pin to 314ffd9, which exposes DaytonaConfig::http_client
and ships on reqwest 0.13.
Drops the three credential-probe unit tests from >1s SLOW to ~0.5s by
skipping macOS proxy discovery on the localhost httpmock requests.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The default OpenAI model moved from gpt-5.5 back to gpt-5.4 in 38b51c4c2,
but this attach test snapshot still asserted gpt-5.5.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Use a lightweight basic probe target for preflight instead of fabricating catalog models, run configured model probes with bounded concurrency, and keep expensive model choices opt-in for defaults and live tests.
Use join_all to fan out provider probes instead of awaiting them
sequentially, and reuse fabro_util::error::collect_chain for the chain
rendering. Carry Provider through ProviderFailure instead of stringifying
it at construction.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`thiserror`-derived `Display` does not walk `#[source]`, so `format!("{err}")`
and `format!("{err:#}")` on a typed error silently produce only the
top-level message — the same format string changes meaning when migrating
from `anyhow::Result` to a typed `Result`. Point at
`fabro_util::error::collect_chain` as the canonical helper and broaden
the test guidance to cover typed errors.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`fabro_llm::Error`'s Display only renders the top-level message field for
`Network`/`Stream`/`Configuration`/`RequestTimeout` variants — the
`#[source]` chain is dropped. Walk the chain at the rendering boundary
so connectivity failures (DNS, connection refused, TLS) surface their
underlying cause in `fabro doctor` output.
Per docs/internal/error-handling-strategy.md, CLI surfaces should render
the full cause chain. Adds a regression test that walks `err.source()`
on a typed Network error with an inner io::Error.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
`fabro doctor` now classifies LLM provider connectivity and auth probe
failures as `CheckStatus::Error` (so the command exits non-zero) and
surfaces the actual probe error text — truncated to one short line per
provider — instead of the generic "Connectivity issues with: <provider>".
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Propagate OpenAI Responses SSE error and response.failed events as structured provider errors, and treat response.incomplete as a normal length finish with partial output preserved.
Also preserve those stream errors through Codex-mode complete_via_stream and add agent coverage proving quota failures do not replay the turn.
GPT-5.5 (released 2026-04-23) replaces 5.4 as the default OpenAI model.
Live integration tests confirm both new IDs respond on the OpenAI API;
they require default temperature like other reasoning models.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PR #202 shipped a new CLI subcommand without regenerating
docs/public/reference/cli.mdx, so the Generated Docs CI job failed on
push. The implement-plan workflow's verify gate had no equivalent of
`cargo dev docs check`.
Append `cargo dev docs refresh && cargo dev docs check` to verify so
the gate auto-fixes drift and surfaces real authoring errors (missing
help text, removed generated-region fences) through the fixup loop.
Also broaden the fixup prompt to cover docs errors.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The implement-plan and smoke workflows ran clippy without --all-targets,
so test, example, and bench targets were skipped. CI runs clippy with
--all-targets, so lint errors in test code passed the workflow's verify
gate but failed CI on push. Aligns the workflow lint commands with CI
and CLAUDE.md.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Use BilledTokenCounts::default() for the non-LLM branch, hoist the
by-model stage count and hasLlmStages predicate out of JSX, and drop
the in-test for-loop in favor of iterator-based assertions.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Include completed stages without LLM usage in run billing responses so command-only runs still show runtime rows. Keep token and model aggregates scoped to billed LLM usage, and render placeholder values in the web billing table.
