feat(*): take the agents page to the agent hub prototype

[0xKT]

Summary

Takes the web UI's agents page (rail entry "Agent Hub", page "Agents") to the prototype's agent hub, and gives the wire what the prototype needed and did not have: a default model on an agent row.

Base: refactor/ui_web_architecture, the long-lived architecture branch (#475), not main. The diff shown here is the diff that merges into that branch.

What the page does now:

• Three sections, never folded: connected, available to connect, not installed on this machine. One control per row (connect / disconnect / retry / install), a status dot only where there is something to say, and a one-line summary from a small catalogue of what each agent is and who makes it.
• A 520x500 sheet in the shared detail host: what the agent is good at (editable, saved on blur), the API key where one is needed, the install command and site for an absent agent with a re-check, a Test button, and the model pill.
• A write in flight is held on the row as "connecting" for its length; a refusal -- a model pick included -- stays on the row as red text with Retry rather than as a toast. A running test keeps a Stop control.

The model pill, by the row's own rule (model_source, a new wire field):

• The built-in row picks from raven's connected providers. The pick is stored naming its provider (stored_model_id, the way config.set model stores the host's), validated against a provider raven knows that holds a usable credential, and paired with that credential on every dispatch through ProviderPool.bind_pin -- the backend resolves the pin per run, so a cached backend never runs one conversation's key against another's model. The write and the dispatch read the provider off the stored id through one function, stored_provider_name (the inverse of stored_model_id), so a section raven has no spec for resolves to that section on both sides.
• An acp row, raven's own shipped products included, picks from the choices its handshake advertised (model_choices) and the value is pushed over session/set_config_option, through one resolver (session_model_for) for a spawn, a direct chat and a DAG node (model_for threaded through the runner).
• An openai or cli row has no menu; the pill is disabled. Unset reads by ownership: one of raven's own "follows the main Raven", a third party runs on "its own default". Ownership is the wire's own (a new field: built-in, vendored, or an acp row whose handshake named Raven), with the catalogue as the fallback for a server that predates it. The picker offers listed ids only; there is no typed entry, since an acp row's write takes exact menu values and a built-in row's adds to no provider.

Deviations from the prototype, deliberately:

• No cancel while connecting; the row shows "connecting" until the server answers.
• Third parties do not pick from raven's providers: there is no path that would deliver raven's credential to them, so they pick from their own advertised menu.
• Raven-PPT is a hidden row and is not listed; "build an agent" is not in this PR.
• The old page's rename control, transport chips and test-cost notes are gone; the prototype has none of them.
• The built-in row's description stays read-only in the sheet (the server now takes the write; the field follows later).

Install data (commands and sites) comes from each vendor's own documentation: Claude Code, Codex, OpenCode, Hermes, OpenClaw, MiroThinker, GitHub Copilot, Qwen Code, CodeBuddy, Qoder, Grok, Kimi Code, Pi.

Wire changes: SubagentRow.own / model / model_choices / model_source; subagents.update takes model, provider, clear_model (clear wins) and description: null (back to the factory text; a built-in row goes back to following its seed); subagents.test takes source: "vendored"; ThirdPartyAcpSubagentConfig.model (and schemas/subagent.schema.json re-exported for it); SubagentManager(provider_pool=...) is injected by the loop and the playbook CLI.

Reviewed in two passes (three read-only reviewers plus a refuter per finding): ten findings confirmed, all fixed in the backend commit -- among them a dispatch that withheld the model from a cached built-in backend, a model_source that promised a write the server refused, and a built-in override edit that would have shadowed a legacy-spelled row.

Review round (gloryfromca, LivXue): all five of the first review's blockers and the second review's two blockers plus its two "fix or file" items where a fix was in reach are addressed in four commits -- the one-resolver read of a stored built-in model (stored_provider_name) with the spec-less section case tested on both sides and an unreadable config raised rather than blamed on the pick, the unusable-pin warning keyed per pin, a refused acp push said once per value instead of once per turn, the wire's own flag read by the page, the stored-prefix spelling matched to the host slug, no typed picks, a Stop for a running test, the model op held on the row, and the stale narrative in the manifest, the CONTRIBUTING ledger rows and two comments. The first reviewer's two later nonblocking follow-ups are in as well: the Claude Code preset no longer requires a claude on PATH (the adapter's SDK pin carries the CLI), and a stored id whose prefix names no provider section is read as part of the id, so a hand-written gateway-served id keeps its pin. The second review's finding 4 is fixed on the record side: a failed manual Test writes its verdict over the previous record's capabilities rather than erasing them, and a row recorded before the menu existed reads "attention, model menu not measured" until a re-verify lands instead of a stored "ready" (CapabilitySnapshot.model_menu_measured). Deferred by the owner as a feature the prototype does not draw: a row field saying whether the stored model is in effect (pin resolved / push accepted), which would be the durable answer to the second review's "silently" in findings 1 and 2. Not in this PR and not tracked as an issue for now.

Four pre-existing defects the live regression surfaced are fixed here as well, each in its own commit:

• raven acp answered session/set_config_option (and session/new) with the process default as the model option's currentValue, because it asked model.options without a session id. A client that renders currentValue as the selected item snapped back to the old model after every switch. Both call sites now ask about the session whose answer it is.
• The acp probe resolved argv[0] alone, and a shim-launched preset's command is an npx one that resolves on any machine with node, so Pi read as connectable wherever pi was not installed and the connect failed a minute later inside the adapter. SHIM_REQUIRED_EXECUTABLES (in presets.py) names the local agent a shim reaches for with its install command; the probe asks after it too and reports the row missing with the install beside it. Only Pi is listed: codex-acp bundles its agent, and claude-agent-acp runs the CLI its @anthropic-ai/claude-agent-sdk pin carries as a per-platform optional dependency (verified in the 0.66.0 package: the SDK's platform package holds the claude binary, and the live run's "Claude Code 2.1.220" is that bundled build, not the machine's), so neither wants an install.
• The Claude Code preset could not connect for an account whose Claude settings name a recent model: claude-agent-acp@0.66.0 bundles CLI 2.1.220 through its SDK pin, and the CLI refuses such a model ("version 2.1.251 or newer is required"). The pin is now 0.79.0 (CLI 2.1.274), measured: handshake with resume, fork and load, six model choices, five modes; connect from the row in 9 s where 0.66.0 failed after 155 s.
• A row connected from the page recorded no capability snapshot, so it read "capabilities not recorded -- run a test", stateless and menuless, until a Test or a restart. The connect now records the handshake the Test would (record_capabilities, shared by both), only when the record is missing, stale or from before the menu.

Type

• Fix
• Feature
• Docs
• CI / tooling
• Refactor
• Other

Verification

Backend (worktree venv with the dev extra):

uv run pytest tests/test_rpc_subagents.py tests/test_rpc_subagents_model.py tests/test_rpc_schema_match.py \
  tests/test_subagent_manager.py tests/test_subagent_probe.py tests/test_subagent_dag_runner.py \
  tests/test_subagent_acp.py tests/test_subagent_builtin_agents.py tests/test_rpc_registration.py \
  tests/test_agent_loop_session_model.py tests/test_agent_loop_subagent_role.py \
  tests/test_subagent_direct_chat.py tests/test_subagent_workdir.py tests/test_agent_loop_web_tools.py \
  tests/test_cli_playbook_commands.py tests/test_config_live.py -q -p no:randomly
1471 passed
uv run pytest tests/test_acp_methods.py tests/test_acp_config_options.py tests/test_subagent_probe.py \
  tests/test_subagent_third_party.py tests/test_rpc_subagents.py -q      # the two fixes: 581 passed
uv run pytest tests/test_rpc_subagents_model.py tests/test_subagent_manager.py tests/test_provider_pool.py \
  tests/test_provider_wire_model.py tests/test_subagent_acp.py tests/test_rpc_subagents.py tests/test_acp_methods.py \
  tests/test_acp_config_options.py tests/test_subagent_probe.py tests/test_subagent_third_party.py -q
1345 passed                              # after the review round
make test-python                         # 24016 passed, 109 skipped (the full unit suite)
make lint-python                         # clean

Web and TUI:

npm run gen:check --prefix ui-web        # generated.ts matches the contract (189 methods)
npm run type-check --prefix ui-web       # clean
npm run lint --prefix ui-web             # 0 errors (5 pre-existing warnings in other domains)
npm test --prefix ui-web                 # 188 files, 2505 tests passed (after the review round)
node ui-web/scripts/check-class-namespace.mjs   # OK, extAgents pinned at 0 / 0
npm run lint:rpc --prefix ui-tui         # in sync
npm run lint:i18n --prefix ui-tui        # up to date

Served page (a test gateway from this branch, an isolated RAVEN_HOME copied from a real one): roster of 5 connected / 4 available / 9 not installed with display names; Raven-Research's pill lists the 55 choices its handshake advertised, a pick lands on the row and reads back from subagents.list; the built-in row's pick is stored as openrouter/anthropic/claude-opus-5 and clears back to null; the three products whose snapshots predated the menu key were re-verified at startup and got their menus.

Live regression on the same gateway: a product row switched off and back on through the row control (the readiness ping relaunches it); MiroThinker connected with a key typed in the sheet; an absent row's install block, copy and re-check ("still not found"); a pick and a clear on Raven-Code's own menu. Then, with the session on DeepSeek and Raven-Research's row set to deepseek/deepseek/deepseek-v4-flash from its menu, one direct-chat turn to a Raven-Research instance: the host sent session/set_config_option {configId: model} with that value (acp frame journal), the child logged Session model: deepseek/deepseek-v4-flash and ran the turn on it (Iteration 1/60 model=deepseek/deepseek-v4-flash), and answered in 15 s. Connecting Claude Code, Codex and Pi from the row each ended in the server's refusal shown red on the row with Retry -- the machine's Claude Code CLI is too old for the adapter's model, Codex is not logged in, Pi's executable is not installed -- so their Test buttons were not reached.

• Relevant tests pass locally
• Relevant lint / type checks pass locally
• User-facing docs or screenshots are updated when needed

Risk

• Behaviour change: the agents page is redrawn; a built-in row's model in config (previously masked by the turn binding on every dispatch) now takes effect, paired with its own credential, and falls back to the conversation's model when no credential can serve it (one warning in the log). An acp row's model is new and off by default.

• The Pi row now reads "not installed" on a machine without the pi executable, where it read "available" before; a hand-written row is not held to this, only a row carrying the preset's provenance.

• An acp row whose capability snapshot predates the model menu reads "attention" with a "run a test" detail until the boot-time re-verify or a manual Test records the menu; before, it read the old "ready" with a disabled model pill. A failed manual Test no longer drops a row's recorded menu and statefulness.

• A built-in row's bare stored id (written by hand before this branch) is now paired by keyword on read and re-stored prefixed on its next write; a pick under a spec-less section is accepted only for an id that section lists. A hand-written id whose prefix names no provider section (deepseek-ai/DeepSeek-V3) keeps its provider derived by the pool -- the configured gateway, as such an id always ran.

• Rollback: revert the PR; the new config fields (model on an acp row, a built-in override row) are ignored by the previous code.

• Security impact considered

• Backward compatibility considered

• Rollback path is clear for risky changes

Related Issues

N/A

[gloryfromca]

Blocking: the agent picker and test sheet still have user-visible paths that cannot behave as presented.

I found four functional blockers and one repository-rule blocker, marked inline. I covered the full merge-base diff, the extAgents callers and prior implementation, model dispatch through spawn/direct-chat/DAG/playbook, generated wire compatibility, commit history, AGENTS.md, CONTEXT-MAP.md, ui-web/CONTEXT.md and CONTRIBUTING.md, and the rewritten tests for removed coverage.

Verification: uv run pytest over the PR's 16 backend suites: 1471 passed; npm test in ui-web: 188 files / 2501 tests passed; focused extAgents suite: 5 files / 86 tests passed; type-check and generated-client check passed; ESLint completed with 0 errors and 5 warnings in untouched domains; class-namespace and large-file checks passed.

[gloryfromca]

Blocking: the five previously reported findings still need to be addressed.

The sole change since my previous review translates the stylesheet comment into English. That fixes the source-language gate failure, but it does not modify the implementation or test files behind any of the five open threads. Direct inspection confirms that all five findings still stand; I found no additional defect in this comment-only delta.

Verification:

• uv run python scripts/check_source_language.py github/refactor/ui_web_architecture...HEAD — passed.
• npm test -- --run src/features/extAgents — 5 files and 86 tests passed.
• npm run type-check — passed.

Coverage: I reviewed the new delta and commit history, rechecked the existing findings against the current implementation and callers, and carried forward the prior full-diff review of project rules, backward compatibility, architecture boundaries, and test integrity. No test was weakened in this revision. The five threads remain open because none is settled.

[LivXue]

Request changes

Thanks -- the page reads well, and the lane work behind the pill is the right shape: one session_model_for for a spawn, a chat and a DAG node, and the acp-only keyword guarded by signature rather than by a kind check. Two findings block, both on the built-in row's model path; the two after them should be fixed or filed. Everything below was read at the head commit; raven/providers/pool.py is unchanged by this PR, so its behaviour is taken as given.

Blocking

1. A built-in row's stored model and the model the dispatch runs can disagree, silently. subagents.update stores a provider-qualified id, and both halves this PR adds read it back by resolving the provider with find_by_model alone -- which fails for an id whose prefix names a section raven carries no spec for. The write side then refuses the picker's own pick with a message blaming the model or the credential, both of which are fine; the bare-id route stores the id, and every dispatch afterwards runs the conversation's pair (or the gateway's key) while the row keeps reporting the stored id. Nothing on the row or the wire says the pin was dropped, and the one log line fires once per resolver, so a later, different unusable pin logs nothing at all. Fix direction: one resolver on both sides (see the inline comments on _host_pair and _row_pin), a refusal that names its actual cause, and ideally a row field for "the pin resolved".

2. An acp model push the agent refuses is retried every turn, and stays on the row as if it landed. _set_model runs on every route into a session, so a permanent refusal is re-attempted as if transient, one warning per attempt with no bound. The row (subagents.list) and the instances report keep showing the pick, and the refusal is a transient toast -- not the held red text with Retry the description promises. Fix direction: the same held-failure treatment the other verbs get, plus a row field (model_push_error, or model_in_effect); failing that, log once per problem rather than once per turn.

Important

3. A config that cannot be read is reported as a credentials problem (raven/rpc/methods/subagents.py:511). The bare except Exception returns None, which the caller renders as "it names no provider raven knows, or that provider has no usable credentials". A PermissionError, a malformed config, a loader failure: none of it is logged, and the user is told their model or key is at fault. An unreadable config is a server-side error, not a field-validation one.

4. The menu backfill can leave a "ready" row with an unmeasured menu, and a failed Test overwrites a good snapshot. _verify_missing_snapshots (raven/agent/subagent/probe.py:579) records only on ready; when the re-verify does not return ready, the old snapshot -- the one with no modelChoices -- survives and _probe_acp reports its stored ready, so the sheet draws a disabled pill and "managed by itself" for an agent that does offer a menu, with an INFO line as the only trace. Same area: _test_acp records unconditionally (probe.py:509), so a manual Test that fails on a flaky machine replaces a good snapshot and costs the row its menu and its statefulness until a later success. has_model_menu already tells "never measured" from "measured, none"; a wire field carrying that distinction would be the durable fix.

Nits

• ui-web/src/features/extAgents/manifest.ts (untouched here) still says the domain's detail-card button is .xaedit and that the class prefix "waits on the CSS step". This PR is that step: five page.css rules gone, prefix renamed, LEGACY_LOCAL.extAgents down to 0.
• ui-web/CONTRIBUTING.md: the LEGACY_BORROWED row still reads "one of the eleven, warn, is SetupSheet.tsx..." and "11 in the shared components", but this PR's ledger edit moved warn off that list (components 11 -> 10, LEGACY_CHROME_EXPR 10 -> 11). That row's totals look stale independently (the two lists hold 23 names; the row says 34) -- worth re-deriving in the same edit.
• ui-web/src/styles/page.css:4143: the specificity comment still cites .sukey input, whose rules this PR deleted (6 occurrences at base, 1 at head, and it is the comment).
• raven/rpc/models.py:150 (own): added to the wire, read once (source.ts:57), consumed by nothing -- ownership is decided client-side by isOwnRow (catalogue.ts:108). Either render from the field or drop it; the server is the one that knows a shipped product was registered as a plain config row. The description's "Unset reads by ownership (own, a new wire field)" describes a read the code does not make.
• ui-web/src/features/extAgents/catalogue.ts:13 and styles.css:17: both point at something a reader of the tree cannot open -- "sources are listed in the PR that added them", and a my_docs/temp/.../proto_app.css path. The prototype reference is useful provenance; the PR pointer is not a source.
• ui-web/src/features/extAgents/store.ts:304: stopTest has no caller.

Process

• The base is refactor/ui_web_architecture (#475, still open). Please say so in the description, or land #475 first -- otherwise the diff a reviewer sees is not the diff that merges.
• Second description/code mismatch, alongside own: "a refusal stays on the row as red text with Retry" does not hold for the model op (finding 2).

What is good here

• The key removal is clean: 17 i18n keys gone (34 leaves with the en/zh pairs), with no dangling reference left in either front end.
• The class-namespace ledger moves are consistent and self-enforcing -- the pin assertion (check-class-namespace.mjs:859) fails the next PR that leaves a pin behind.
• The lane unification is the right shape, and the acp-only keyword travels only to a run that declares it or takes **kwargs, so a backend written against the older paper is not handed it.
• The contract artifacts (openrpc.json, both generated.ts, the JSON schema) move in the same change as the models they describe.
• The verification section shows the acp path exercised end to end against a live gateway. The built-in path is where it stops short: the pick recorded there is openrouter/..., whose prefix find_by_model resolves, so the spec-less section looks like the case neither the tests nor the live run covered.

[gloryfromca]

Blocking: the five previously reported findings remain unaddressed.

Neither new commit modifies the implementation or test files behind my five open threads, and direct inspection confirms that each still stands. I found no additional blocker in the ACP session-model or shim-probe delta under the fourth-round bar.

Named nonblocking follow-up: SHIM_REQUIRED_EXECUTABLES now requires a PATH-level claude for the Claude Code preset, but the pinned @agentclientprotocol/claude-agent-acp@0.66.0 resolves its native Claude CLI from the platform-specific optional dependency of @anthropic-ai/claude-agent-sdk. A Node-only installation can therefore run the adapter while this probe reports it missing. This change introduces the false negative and it is reachable in ordinary use, but the displayed global-install command is a working escape hatch, so it does not meet the fourth-round blocker threshold.

Verification:

• uv run pytest tests/test_acp_config_options.py tests/test_acp_methods.py tests/test_subagent_probe.py tests/test_subagent_third_party.py -x: 475 passed on Python 3.12.3.
• npm test -- --run src/features/extAgents: 5 files and 86 tests passed.
• npm run type-check: passed.
• Source-language, large-file, and git diff --check gates: passed. The environment has no make, so I ran the large-file target command from the Makefile directly.

Coverage: reviewed the full revision diff and the two-commit delta, surrounding callers and history, AGENTS.md / CLAUDE.md / CONTEXT*.md rules, backward compatibility, Runtime/Web UI architecture boundaries, and whether tests were weakened. The new tests add coverage rather than weakening existing assertions. Existing threads remain open because none is settled.

[gloryfromca]

Blocking: add the required file-level purpose documentation to ui-web/src/features/extAgents/catalogue.test.ts.

Four of my five original findings are fixed; I replied to and resolved those threads. The remaining new test module still begins directly with imports, so it does not satisfy AGENTS.md section 1.1. At this fifth-round bar: the PR introduces the file, the violation is unconditional rather than dependent on a rare runtime combination, and the repository explicitly treats this as a hard rejection rule with no compliant workaround other than adding the file-level documentation.

Named nonblocking follow-ups:

• The previously reported Claude adapter probe false-negative remains unchanged.
• The new stored_provider_name resolver can reinterpret a pre-existing built-in override such as deepseek-ai/DeepSeek-V3, which previously ran through the configured gateway, as an explicit deepseek_ai provider and drop the pin. Reselecting the model under the gateway stores openrouter/deepseek-ai/DeepSeek-V3, so operators have a working recovery and this does not meet the late-round blocker threshold.

Verification:

• uv run pytest tests/test_provider_pool.py tests/test_provider_wire_model.py tests/test_rpc_subagents_model.py tests/test_subagent_acp.py tests/test_subagent_manager.py -x: 764 passed.
• npm test: 188 files and 2505 tests passed.
• npm run type-check and npm run gen:check: passed.
• npm run lint: passed with 0 errors and 5 pre-existing warnings on untouched files.
• Source-language, large-file, class-namespace, and git diff --check gates: passed.

Coverage: reviewed the full target diff and this four-commit delta, surrounding callers and history, AGENTS.md / CLAUDE.md / CONTEXT*.md rules, backward compatibility, Runtime/Web UI architecture boundaries, and test integrity. The added tests strengthen coverage; no test was weakened.

[gloryfromca]

Blocking: add the required file-level purpose documentation to ui-web/src/features/extAgents/catalogue.test.ts.

The new delta correctly stops treating a globally installed claude executable as a prerequisite for claude-agent-acp, and its regression test matches the adapter packaging contract. The remaining blocker is unchanged: this PR-added test module still begins directly with imports, contrary to the mandatory file-level purpose documentation in AGENTS.md section 1.1. At this sixth round it meets all three blocking criteria: this PR introduced the file, every normal checkout contains the violation, and there is no compliant workaround without changing the file.

The previously named legacy configured-gateway model-id compatibility case remains a nonblocking follow-up.

Verification: uv run pytest tests/test_subagent_third_party.py tests/test_subagent_probe.py -x (293 passed); source-language checker passed; large-file checker passed via its underlying uv command because make is unavailable; git diff --check github/refactor/ui_web_architecture...HEAD passed. I reviewed the revision delta and target diff context, affected callers and history, project rules and domain context, backward compatibility, architecture boundaries, and whether tests were weakened.

[0xKT]

Review round, status by finding.

LivXue:

• Blocking 1 (built-in stored model vs dispatch): fixed in 35bf0f2 and 79196f4. One resolver (providers.wire.stored_provider_name) on the write and the dispatch side, the spec-less section tested on both, the unreadable-config refusal removed, the unusable-pin warning keyed per pin, and a prefix that names no section left to the pool to derive (the gateway case).
• Blocking 2 (refused acp push): the write refusal is held on the row with Retry (4749393); the per-turn re-attempt is said once per value (18994e3). A row field for "in effect" is not in this PR -- see Important 4.
• Important 3 (unreadable config reported as a credentials problem): fixed in 35bf0f2.
• Important 4 (menu backfill and failed Test overwriting a snapshot): fixed in 79196f4 and 5bc8c6a. A failed manual Test records its verdict over the previous record's capabilities instead of erasing them (probe.py _test_record), and a row recorded before the menu existed reads attention with "model menu has not been measured yet" until a re-verify lands, instead of the stored ready (CapabilitySnapshot.model_menu_measured). The in-effect row field (findings 1 and 2) stays a proposed follow-up.
• Nits: manifest text, CONTRIBUTING ledger rows (re-derived from the tool: 125 / 15 / 23, warn on the expression list), page.css comment, the two pointers, stopTest's caller -- 5c70ce8 and 4749393.
• Process: the description states the base branch in its first paragraph; the model-op sentence now matches the code.

gloryfromca:

• The five first-round findings: four resolved by you; the catalogue test module's file-level comment is in 79196f4.
• Nonblocking, Claude adapter probe: fixed in ac97e89 (claude_code is out of SHIM_REQUIRED_EXECUTABLES; the SDK pin carries the CLI).
• Nonblocking, legacy gateway-served id: fixed in 79196f4 (stored_provider_name takes the config's provider table; a prefix naming no section is part of the id and the pin goes to the pool's gateway branch, with tests on both sides).

Verification for the round: uv run pytest over the touched backend files: 1345 passed before the follow-ups, then 667 / 607 / 365 passed on the files each follow-up touches (test_subagent_acp.py, test_subagent_probe.py, test_rpc_subagents.py, test_subagent_manager.py, test_provider_wire_model.py, test_rpc_subagents_model.py, test_provider_pool.py); npm test in ui-web, 188 files / 2505 tests; type-check, lint (0 errors), gen:check, check-class-namespace and the TUI catalogue check all pass.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

The file-level purpose documentation now satisfies the remaining repository-rule blocker. This revision also closes both follow-ups I had named: the bundled Claude adapter is no longer rejected for lacking a global executable, and legacy gateway-served model ids remain on the pool derivation path while configured spec-less provider sections resolve explicitly. The snapshot updates preserve prior capabilities across a failed test without collapsing an unmeasured model menu into a measured-empty one. I found no new plain error in the delta, and every thread I opened is now resolved.

Verification: uv run pytest tests/test_provider_wire_model.py tests/test_provider_pool.py tests/test_rpc_subagents.py tests/test_rpc_subagents_model.py tests/test_subagent_acp.py tests/test_subagent_manager.py tests/test_subagent_probe.py -x (926 passed); npm test -- --run src/features/extAgents/catalogue.test.ts (6 passed); source-language, large-file, and git diff --check gates passed. This pass covered the revision delta and full target-diff context from the prior rounds, affected callers and history, project rules and domain context, backward compatibility, architecture boundaries, and test integrity.

[0xKT]

Owner's decision on the one open item: the row field for whether a stored model is in effect (LivXue findings 1 and 2, the raven_loop.py:340 thread) is a feature the prototype does not draw, so it is deferred -- not in this PR and not tracked as an issue for now. Every other finding in the round is fixed or answered on its thread.

[0xKT]

Two more pre-existing defects found by the real-machine acceptance on this branch, fixed in faefa84 and 3d34c9d: the Claude Code preset's adapter pin moves from 0.66.0 to 0.79.0 (0.66.0 bundles CLI 2.1.220, which refuses any recent Claude model with "version 2.1.251 or newer is required", so the row could not connect for such an account; 0.79.0 measured: handshake with resume/fork/load, six model choices, five modes, connect from the row in 9 s, a tool-using turn rendered with the same _meta.claudeCode.toolName / rawOutput / fence shapes the dialect reads), and a row connected from the page now records its capability snapshot at connect (record_capabilities, shared with the manual Test; only when the record is missing, stale or from before the menu), where before it stayed stateless and menuless until a Test or a restart. Tests: test_rpc_subagents.py test_add_records_the_capabilities_of_an_acp_row_that_answered and siblings, test_subagent_acp.py test_the_connect_records_what_a_test_would.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

The acceptance-test delta is sound. The Claude ACP preset now pins a published adapter whose SDK dependency matches the documented compatibility rationale, while retaining the measured dialect shapes. The add/enable path records capabilities only when the snapshot is missing, stale, or predates model menus; it shares the manual-Test writer, treats recording as best effort after a successful ping, and leaves concurrent config changes detectable through the existing fingerprint. I found no new plain error.

Verification: uv run pytest tests/test_rpc_subagents.py tests/test_subagent_acp.py tests/test_subagent_probe.py tests/test_subagent_third_party.py tests/test_acp_dialects.py -x (665 passed); Ruff passed on all touched Python files; source-language, large-file, and git diff --check gates passed. I also verified the npm package metadata for @agentclientprotocol/claude-agent-acp@0.79.0. This pass covered the new delta plus the previously reviewed target-diff context, affected callers and history, project rules and domain context, backward compatibility, architecture boundaries, and test integrity.