Summary

Display model: "<agentId>-acp", provider: "acpx" (the ACP-runtime sentinel) for ACP-keyed sessions in the openclaw sessions listing, instead of the agent's configured model which was misleading. Closes catalog #20.

Bug detail

Every ACP copilot session in openclaw sessions listing reported model: "gpt-5.3-codex", modelProvider: "microsoft-foundry" — but those are the copilot agent's configured model (used for non-ACP / openclaw-agent-driven flows). When the agent runs as an ACP child, copilot CLI selects the actual model internally. The listing was reporting a model that wasn't actually used for those sessions, causing operator confusion.

How to reproduce

Live repro before fix:

```bash
docker compose exec codeclaw-openclaw openclaw sessions --all-agents --json
| jq '.sessions[] | select(.key | contains("acp:")) | {key, model, modelProvider}'

Before: all ACP sessions show "model": "gpt-5.3-codex", "modelProvider": "microsoft-foundry"

After: all ACP sessions show "model": "copilot-acp", "modelProvider": "acpx"

```

Unit-level:

```bash
git checkout 5f12c52 # red-test cherry-pick
pnpm test src/commands/sessions.acp-model-display.test.ts

Before fix: 2 RED (model is the agent default; should be copilot-acp)

After fix commit e744c1a + cleanup 87b4eb7: all 3 GREEN

```

Root cause

resolveSessionDisplayModelRef (src/commands/sessions-display-model.ts:123-148) had zero ACP-awareness. It returned the agent-default model regardless of session-key shape — and the resolver is intentionally pure (per its existing design contract; we don't want to teach it about ACP).

Fix approach

Sentinel overlay at the call sites — keeps the resolver pure. Mirrors the established #18 fix pattern:

1. Added private applyAcpModelOverlayIfNeeded helper in src/commands/sessions.ts that, when isAcpSessionKey(row.key) is true, replaces the resolver's output with { provider: "acpx", model: "<agentId>-acp" }. The agentId is extracted via parseAgentSessionKey, so multi-agent ACP works (copilot-acp, codex-acp, etc.).
2. Applied the overlay at three emit sites: the allRows construction path, the JSON --json emit path, and the table-display per-row path.
3. The 5 other "parallel" files in the Task's allowed list (gateway server-methods sessions, gateway session-utils, doctor-claude-cli, agents-list-tool, status.summary.runtime) were verified NOT to call resolveSessionDisplayModelRef and so don't need the overlay. They use a different resolver (resolveSessionDisplayModelIdentityRef) which is a separate surface tracked elsewhere.

The helper currently lives inline in sessions.ts because the related PR for catalog #18 introduces a shared src/agents/acp-runtime-overlay.ts that would be the natural home. A // TODO comment in the helper marks the planned migration once both PRs land.

Tests

• Red test (now GREEN): src/commands/sessions.acp-model-display.test.ts — 3 cases (2 RED → GREEN, 1 GREEN control stays GREEN).
• Adjacent suites (4 files): sessions.model-resolution.test.ts, sessions.default-agent-store.test.ts, sessions.test.ts, sessions.acp-model-display.test.ts — 25/25 pass after the bridge-session regression case.
• Cleanup commit: removes a dead import that the fix replaced; oxlint now clean.

Real behavior proof

• Behavior or issue addressed: ACP control-plane sessions in openclaw sessions --all-agents --json should display the ACP runtime sentinel instead of the agent's configured Copilot model/provider.
• Real environment tested: Crabbox Azure lease blue-crayfish / cbx_c22caa546489, with the after-fix capture saved in .crabbox/captures/cbx_c22caa546489-20260513T201301Z.tar.gz and posted by a maintainer at fix(sessions): display ACP-runtime model sentinel for ACP-keyed sessions #79543 (comment).
• Exact steps or command run after this patch: Ran the focused sessions command path in Crabbox and inspected openclaw sessions --all-agents --json rows for an ACP control-plane session and a direct control session; also ran pnpm test src/commands/sessions.acp-model-display.test.ts src/commands/sessions.model-resolution.test.ts src/commands/sessions.default-agent-store.test.ts src/commands/sessions.test.ts and touched-file oxlint.
• Evidence after fix: Terminal output from the Crabbox stdout capture:

AFTER_FIX_ACP_ROW={"key":"agent:copilot:acp:<redacted>","model":"copilot-acp","modelProvider":"acpx"}
CONTROL_DIRECT_ROW={"key":"agent:copilot:main","model":"gpt-5.5","modelProvider":"github-copilot"}
Test Files 4 passed (4)
Tests 25 passed (25)
Found 0 warnings and 0 errors.

• Observed result after fix: The ACP row reports model: "copilot-acp" and modelProvider: "acpx", while the direct control row still reports model: "gpt-5.5" and modelProvider: "github-copilot".
• What was not tested: No additional gaps for this sessions display path; exact-head GitHub CI is still running after the maintainer rebase.

Catalog reference

Catalog finding: #20. Investigation lives in branch edpiva/acp-native-session-investigation (not yet upstream) at docs/plans/2026-05-08-acp-findings-catalog.md.

Verification commands

```bash
pnpm test src/commands/sessions.acp-model-display.test.ts # 3/3 GREEN
pnpm test src/commands/sessions.model-resolution.test.ts
src/commands/sessions.default-agent-store.test.ts
src/commands/sessions.test.ts
src/commands/sessions.acp-model-display.test.ts # 24/24 GREEN
node scripts/run-oxlint.mjs src/commands/sessions.ts # 0 errors
```

Notes for reviewers

• Sentinel choice: model: "<agentId>-acp" (e.g., copilot-acp, codex-acp) per the catalog spec. Open to changing the format if you'd prefer something different (e.g., acpx:<agentId>).
• The applyAcpModelOverlayIfNeeded helper is intentionally inlined in sessions.ts until the parallel PR for catalog RPC agent configuration not working #18 lands its src/agents/acp-runtime-overlay.ts helper file. Then a small refactor PR can move the helper there.
• The resolver resolveSessionDisplayModelRef is intentionally unchanged — the overlay is at call sites, preserving the resolver's purity.

Live behavior repro (deployed container)

Container: codeclaw-openclaw:clean running OpenClaw 2026.5.6 (e9d4a0a)
PII note: UUIDs are random-not-PII; no PII to redact in this slice.

Every ACP session reports the agent-config model, not the ACP runtime sentinel

$ docker exec codeclaw-openclaw openclaw sessions --all-agents --json \
    | jq -c '.sessions[] | select(.key | startswith("agent:copilot:acp:")) | {key, model, modelProvider}'

{"key":"agent:copilot:acp:86b7b5af-…","model":"gpt-5.5","modelProvider":"github-copilot"}
{"key":"agent:copilot:acp:d6fb7506-…","model":"gpt-5.5","modelProvider":"github-copilot"}
{"key":"agent:copilot:acp:f5ab8dba-…","model":"gpt-5.5","modelProvider":"github-copilot"}
{"key":"agent:copilot:acp:1f854fa2-…","model":"gpt-5.5","modelProvider":"github-copilot"}
{"key":"agent:copilot:acp:930316f7-…","model":"gpt-5.5","modelProvider":"github-copilot"}
{"key":"agent:copilot:acp:63867cf7-…","model":"gpt-5.5","modelProvider":"github-copilot"}
{"key":"agent:copilot:acp:8fb42bab-…","model":"gpt-5.5","modelProvider":"github-copilot"}

gpt-5.5 / github-copilot is the copilot agent's configured model (used for non-ACP / openclaw-agent-driven flows). Under ACP the copilot CLI selects its own model internally; the listing was misreporting a model that did not run.

After this PR, the same query returns model: "copilot-acp", modelProvider: "acpx" for those rows — the actual ACP runtime sentinel.