| name | cl_surfacing-agent-work |
| description | Use when building or changing any UI/feature that shows an agent's progress, output, or "what it's doing" (live or historical), when choosing how to stream agent output, or when a user says an agent view "looks stuck" / "I can't see the output". Prevents designing an agent viewer before knowing what the agent actually produces — and promising a rich timeline an agent can't deliver. Skip if the view is blank/stuck because the session-log path or resolution is wrong (CLAUDE_CONFIG_DIR/CODEX_HOME, missing JSONL) → cl_locate-agent-session-logs; this skill is for designing what the viewer shows and how output streams. Skip for non-agent UI styling → cl_ui-design-philosophy, and for generic "only updates after refresh" staleness → cl_live-state-sync. |
Surfacing an Agent's Work (without overpromising)
The expensive, repeated mistake on this codebase: building the viewer before
grounding in the agent's real execution model. A polished "watch the agent
think / read / work" surface is worthless — and reads as stuck — if the agent
behind it is a single completion that produces no such trace. Match the surface to
what the agent actually emits, and tell the user the limits before you build.
Step 0 — classify the agent BEFORE designing anything
Find the spawn and answer one question: is this an agentic loop or a one-shot
completion?
| Shape | How to spot it | What it can show |
|---|
| Agentic loop | spawned to read/iterate with tools: the heal agent, and PRD-summary + coverage-mapping (their prompts list file paths and make the agent read the docs/specs itself) — long-running, writes tool_use / tool_result / thinking to its session log | A genuine timeline — reads, thinks, tool calls, results, multiple turns |
| One-shot completion | codex exec, or a prompt that inlines all context so the model answers in one turn (the evaluation rewrite, eval judge) — seconds, no tool calls | Only the prompt + one final answer. No reads. No thinking trace. No tool steps. |
Don't classify by spawn flags alone. claude -p --dangerously-skip-permissions
already has tools ON — what makes a spawn one-shot in practice is a prompt that
inlines every input, so the model never needs to call a tool. PRD-summary +
coverage-mapping were one-shot until their prompts were changed (June 2026) to list
file paths and require reading them — now they stream a real read/think/tool
timeline through AgentSessionView. The evaluation rewrite still inlines its evidence
packet, so it stays one-shot.
If it's one-shot, say so up front and scope the view to reality. Do NOT mount a
tool-loop timeline on it and hope — you'll ship "1 event, looks frozen", the user
will bounce, and you'll rebuild it (this happened 3× before PRD/coverage were made
agentic). The durable fix for a one-shot you want to watch is to make it agentic
(stop inlining; hand it paths to read), not to dress up the empty timeline.
Two transports — pick the one that fits the shape
| Transport | What it is | Good for | Cannot |
|---|
On-disk session JSONL — AgentSessionView tails it (REST snapshot + /ws/.../agent-session, parsed by agent-session-log.ts) | The agent CLI's own session file; the parser emits complete events only | Agentic loops; historical replay; the structured rail (thinking/tool/result rows, model+session header) | Token-stream a one-shot — the assistant block only lands at completion, by which point the job is done and the view moves on, so you see just the prompt |
stdout --output-format=stream-json --include-partial-messages | The live token stream on stdout; parse deltas yourself (see features/agent-sessions/logic/agent-stream.ts) | Watching a one-shot write its answer in real time | Give you the structured tool/think rows for free — you render the text |
Rule of thumb: agentic loop → AgentSessionView (file tail). One-shot you want to
watch → stream-json stdout → a live log. They are not interchangeable; the file
tail physically cannot show token partials (the JSONL records whole blocks, and the
shared parser drops partial lines). See [[cl_ui-design-philosophy]] "One agent
timeline everywhere" for when the structured rail IS the right call.
Honesty / anti-overpromise (the rule that would have saved 3 rounds)
- When a user asks to "see what the agent is doing," answer with the agent's shape
first, then build. "These are one-shot inferences — there's no read/think trace;
the most you can watch is the answer stream in" is the honest framing. Saying it at
design time is cheap; discovering it after shipping is three rebuilds.
- Don't ship a thinner proxy (a flat
<pre> log behind a toggle) for the thing the
user explicitly asked to see and call it done. Either deliver the real view or
surface the constraint and decide together. A toggle that flips between two equally
empty states "does nothing".
- Verify the data exists before designing the view that displays it. If you can't
confirm the wire shape locally (e.g. claude's stream-json — needs the user's
canary-apply env), say so, parse defensively, and degrade safely.
Persist the LIVE lifecycle, not just navigation
A generation/heal job is durable server-side. If the UI shows it, the UI must
rehydrate the running job on open (query the job list, re-attach the poller) —
not just restore which view/feature was selected. Persisting navigation but not the
in-flight job is why "refresh during generating loses the state" even though the
header still says Generating. One owner for the job lifecycle; every panel reads
it (see [[cl_ui-design-philosophy]] "One owner for a long-lived lifecycle" and
[[cl_async-task-ux]]).
Common mistakes (all real, all from one feature)
| Mistake | Symptom | Fix |
|---|
| Mounted a tool-loop timeline on a one-shot agent | "1 event", looks stuck | Classify first; stream-json stdout → live log for one-shots |
| Promised "watch it think/read" for an agent with no tools | User bounces twice | State the shape up front; scope the view to it |
| File-tail viewer expected to token-stream | Only the prompt shows until completion, then the view closes | Use stdout stream-json for the live token feel |
| Persisted view+feature but not the running job | Refresh mid-job → empty ledger | Rehydrate the active job on mount |
| Deferred the core "see the agent" ask with a flat-log placeholder | "the toggle does nothing" | Do it properly or surface the real constraint |
Verify
Stream parsing is pure + unit-tested (features/agent-sessions/logic/agent-stream.test.ts) so the logic
is checkable without a live agent; the actual wire shape is the user's canary-apply
trial (never run it — see cl_verify-changes). Component behaviour is happy-dom-tested.
