fix(workspace): remove "<agent> doesn't report steps" from the Work card (#3001)

[dolho]

Summary

The Workspace Work card no longer says " doesn't report steps", at any point in a run.

• The sentence was ruled for ent#525, before fix(agent-runtime): kill npx MCP orphans outside claude pgid that hold stdout pipe open (#618) #620 put the agent's live activity line on the same card. Beside "Thinking" / "Reading …" it read as the card contradicting itself, and before the first line it read as a fault. Ruled 2026-10-01: remove it outright.
• A running card now shows the live activity line, or the stages when the agent publishes them, or nothing.
• "Steps could not be read right now." stays. It reports a real failure: the agent is stopped or unreachable, or two runs are on one agent.

Supersedes the closed #3002. The first commit here re-landed #3002's narrower rule (hide the sentence once activity has shown) on top of #2972. The second commit replaces it with full removal.

Changes

• portalWork.js: stepsLine(steps) returns { kind: 'none', text: '' } for an agent with no stages. The activitySeen option and the name/claim split (who), which existed only for the sentence, are gone.
• PortalWorkCard.vue: none is silent like pending.
  • outside the chat: no steps line;
  • in the chat's one-line card (reserveLiveRows, bug(workspace): the live chat card jumps ~2s into a turn — "<agent> doesn't report steps." and Stop pop in after first paint #2964): the reserved row stays, blank, aria-hidden and h-4, so the card never changes shape.
• Tests:
  • workStepsActivity.spec.js (new, mounted, Safety-critical UI logic keeps landing in the one tier with no executable coverage — and the stated reason is false #2918);
  • four existing bug(workspace): the live chat card jumps ~2s into a turn — "<agent> doesn't report steps." and Stop pop in after first paint #2964 / ent#525 tests that asserted the sentence now assert its absence, keeping the shape guarantees they protect.
• Docs: architecture/workspace.md, feature-flows/workspace-work.md, requirements/core-agent.md, user-docs/operations/executions.md.

Test Plan

• npx vitest run tests/unit/workStepsActivity.spec.js covers:
  • no sentence before the first activity line, beside a live line, or between heartbeats;
  • "could not be read" still shown;
  • the chat card's row blank, aria-hidden and h-4.
• Mutation: putting the sentence back in stepsLine turns the spec red. The card's own tests stay green because the card silences none by its kind, so there are two independent safeguards.
• Full frontend suite: 244 files, 4,484 tests. The design-token check passes.
• Enterprise-docs guard over the doc changes: 0 hits.

Fixes #3001

🤖 Generated with Claude Code

[dolho]

/review report: fix/3001-work-steps-activity-v2 → dev

Files changed: 9 (+137/-54). Scope: CLEAN. Plan completion: 2/2 done (remove the sentence everywhere; mounted test).
Merges cleanly with today's dev (git merge-tree), and also together with #3155, #3156, #3157 and #3158 (octopus merge).

Execution coverage

changed symbol / test	executed by	live consumer	verdict
stepsLine(steps), kind none → empty text	workStepsActivity.spec.js, portalWork.spec.js	PortalWorkCard.vue steps computed, its only caller	✅ executed
stepsSilent (pending ∪ none)	workStepsActivity.spec.js (mounted), portalWorkCardLiveShape.spec.js	the card's steps row	✅ executed
removed who / portal-work-sentence-*	3b in the live-shape spec asserts they're absent	none left on dev (grepped)	✅

Fix mutation: putting the sentence back in stepsLine turns the spec red.

Live check (local dev stack, 2026-10-01, dev + this PR)

A real turn on proj-beta in a Workspace chat, sampled every 500 ms: 42 samples, 4 with the live activity line, 9 with the reserved row blank, 0 containing "doesn't report steps".

Critical findings

None.

Informational

None.

Clean categories

• SQL/auth/credentials: no backend change.
• Enum completeness: kind values pending|unknown|none|stages; every consumer is in PortalWorkCard.vue, and all were checked.
• Docs: the 4 docs describing the sentence are updated; enterprise-docs guard: 0 hits.
• Frontend: tokens only, the row footprint is kept (principle 4), mounted tests (Safety-critical UI logic keeps landing in the one tier with no executable coverage — and the stated reason is false #2918).

Summary: Critical 0 · Informational 0 · Scope clean.

[vybe]

merge-train: pushed b4d3cd315. It changes doc and comment text only.

docs/memory/feature-flows/workspace-work.md still described the removed sentence as live: Ruling 2's "says so", "only the agent's name truncates", and "never 'doesn't report steps'". So did comments in PortalWorkCard.vue, client_portal/work/models.py and pipeline_state.py. They now match none-is-silent. There is no behaviour change.

The PR removes the sentence in every case, citing "Ruled 2026-10-01". #3001's Expected section had kept it for agents that report nothing. A one-line note on #3001 recording that ruling would help the next reader.

[vybe]

merge-train: batch validated on train PR #3179 (all checks green).