stage-11-agentics/c11/c11-workspace-review

c11 Workspace Review

A focused audit of one workspace — every pane, every agent inside it — boiled down to one clear next action for the operator.

This is c11-full-review scoped down. The principles are identical: trust the terminal scrollback over sidebar metadata, distinguish blocked on the operator from progressing on its own, pick one top call. The only difference is that the survey is narrowed to the workspace the operator is currently in.

When to invoke

Operator phrases that match: "what's going on in this workspace", "what should I do next here", "triage this project", "where are we in [project]".

Not for: implementing tickets, reviewing code, planning new work.

Procedure

1. Identify the workspace and its panes

c11 identify --json    # caller's workspace_ref
c11 tree --json        # scoped to current workspace by default

Build the map: which panes, which surfaces (terminals, browsers, markdown), what each tab is titled.

2. Read every non-trivial terminal pane in parallel

c11 read-screen --workspace <ws> --surface <surf> --scrollback --lines 200

Issue all reads in one tool batch. They are independent.

For each transcript, extract:

• What is the agent actually doing now? (Last 1–3 interactions trump the tab title.)
• Has it finished? "Merged", PR links, final summary, idle prompt.
• Is it waiting on the operator? Clarifying questions, "want me to…", "let me know how to proceed".
• Is it stuck? Errors, retries, ambiguous state.
• Cross-references within the workspace — pane A's plan may depend on pane B's output.

The sidebar lies. The terminal usually doesn't. If sidebar says "running" but the transcript shows the agent posted its summary and went idle, trust the transcript.

3. Use external tools only when the terminal is ambiguous

The transcript is usually enough. Reach for gh pr list, the forgejo MCP, lattice ticket show, etc. only when:

• A transcript says "PR opened" without confirming merge state, AND that matters for ranking.
• Two panes contradict each other.
• An external claim ("CI red", "deployed", "all green") needs a real check before the operator acts on it.

4. Synthesize

For each pane, place it in one of:

• Blocked on operator — needs a decision, answer, or "go". Highest-leverage bucket.
• In flight (verified) — actively progressing, no operator input needed. Name the most recent concrete signal.
• Releasable — finished, idle, or placeholder. Worth closing.
• Defer — exists but isn't load-bearing right now.

Then pick one top call — the single action that, if the operator did nothing else, would unblock the most downstream work in this workspace.

5. Report

Free-form prose. Plain English.

Required sections, in this order:

• Top call — one sentence. The single most important next action in this workspace.
• Blocked on you — bullets, each: which agent (de-jargoned + surface ref), what decision it needs, why it matters. Say "nothing" if empty.
• In flight — bullets, each: agent, what it is actually doing (from transcript), most recent concrete signal.
• Releasable — what can be closed/released. Say "nothing" if empty.

If something in this workspace is blocked on another workspace (e.g., a delegator here waiting on a decision the operator owes an agent in workspace 2), surface it explicitly — that's the most common cross-workspace dependency.

De-jargon every agent name on first use. "GR Architect" → "GR Architect (Gate-Rerun architect, surface:10)".

Worked example

Operator is in workspace 2 (holodeck). After `c11 tree --json` and reading 4 panes in parallel:

• Pane A: GR Orchestrator, last line "PRs #29, #30, #31, #32, #33 all merged ✓" 12 minutes ago, then idle.
• Pane B: GR Architect, last line "Phase 2 handoff complete. Idle."
• Pane C: HOLO-59 delegator, last line "PR #31 merged. Standing by."
• Pane D: HOLO-38 architect, no recent activity, last interaction 2 hours ago — "awaiting decision on rerun scope".

Good output:

Top call: Decide HOLO-38 rerun scope — the HOLO-38 architect (surface:14) has been stuck on that for two hours, and the gate-rerun PRs it depends on already merged, so this is the only thing holding the next phase.

Blocked on you

- HOLO-38 architect (surface:14) — asked "rerun scope?" 2 hours ago, no response. Now actionable because dependency PRs landed.

- GR Architect (Gate-Rerun architect, surface:10) — Phase 2 done, idle. Needs "prep HOLO-38" or "stand down".

In flight

- Nothing actively progressing. The gate-rerun campaign landed; everyone is either idle-waiting or blocked on you.

Releasable

- HOLO-59 delegator (surface:13) and the other 4 delegator panes (surface:15–19) — PRs merged, agents standing by. Safe to close.

Notice what made it good: the top call names the bottleneck precisely (2 hours stuck, dependency just unblocked); the in-flight section is honest about "nothing"; the releasable items name why they are stale.

Anti-patterns

• Five generic "Check / Monitor / Review" items. Pre-digest into decisions, not chores.
• Trusting sidebar metadata over transcript content.
• "Investigate why X is throttled" without first reading the transcript to see if "throttled" actually means "finished and idle".

Related

• c11-full-review — same procedure across every workspace in the current window.