[trissim]

Domain tool declarations imported ACP's ToolKind solely to name an action, and native image records lived inside the ACP prompt decoder. Every CLI paid for the schema through those vocabulary imports.

Tool actions are now plain ABC capabilities registered by ToolAction.init_subclass (tool_declaration.py:38-46), composed into tools through their bases. ToolDeclaration.declared_action() reads the C3 MRO, defaulting to OtherAction; FamilyMeta is unchanged. Live ToolStart.kind and the derived ToolTranscript.kind carry the capability class. One AcpToolAction MroDispatch in agent_event_updates.py owns the five SDK literal spellings, used by Core publication and Toad's saved-call reconstruction.

ImageInput and its byte/MIME validation live in image_input.py. Every Core and paired Toad importer moved; image_inputs.py only owns ACP projection and does not re-export the record. Existing tests now expect capability classes for domain values and keep their SDK-string assertions at the ACP output.

Paired Toad dependency

OpenHCSDev/toad#1165 at 5727cccb9 migrates the three image imports and both recorded-tool readers. Core #2913 must land first, then the Toad ratchet must rerun: its current failure at run 40025551025 is TypeErrors +1 in prompt.py, clipboard_image.py and prompt/resource.py because Core main lacks the new image_input module. Toad's workflow resolves Core from its declared source at .github/workflows/ci.yml:42-51. Against this Core checkout, all paired modules import and a saved ReadAction becomes SDK kind read. No compatibility export or CI exception is added.

Storage and source evidence

ToolStart is an in-process observation (agent_events.py:1-5), created at pi_events.py:924 and dispatched through turn_progress.py:269-278 to ACP publication. The stored transcript has tool_name; its action is a property, not a stored field (transcript_events.py:422-423). A sample encoded ToolStartTranscript contains kind=tool_start, tool_call_id, tool_name and raw_input, with no stored action kind. Thus no stored spelling changes.

Existing refactor-audit Package.load parsed 450 production modules and 486 test modules before, and 451 production modules after, with zero parse omissions. The paired Toad scan parsed 288 modules with zero omissions. The inventory covered declarations, action assignments/readers, constructors, annotated image fields and imports. Dynamic dispatch was followed through the existing MroDispatch consumers; the interpreter checked all seven native tool declarations and paired Toad conversion. The independent Codex wire ImageInput declaration is unchanged.

After AST inspection: tool_declaration, agent_events, native_tools, image_input and pi_commands each have 0 absolute acp SDK imports. Remaining production import from image_inputs is acp.py:75 importing prompt_images. acp_kind has 0 hits in src/tests/tools.

The requested literal grep is printed honestly; it still matches the internal ACP failure module, not an SDK import:

src/agent_comms/agent_events.py:14:from .acp_failure import (

No route-lock files are changed.

Importtime at the delivered head

Base 01aa67d versus ff66279, pinned installed core-a02979325 runtime, nice 19, PYTHONPATH set to the checked-out src. One before/after pair per entry, measured consecutively. These are TOTAL lines in milliseconds; elapsed totals vary with machine load. ACP still loads through the reserved controller/CLI dependencies; this change removes the vocabulary edges rather than completing the entire startup target.

Entry	Before TOTAL ms	After TOTAL ms
agent_comms.cli	1821.949	1599.255
agent_comms.commander.cli	1797.037	1686.35
agent_comms.cli_bridges.claude	2103.284	1374.855

agent_comms.cli

acp(311) <- acp.exceptions(311) <- agent_comms.cli_commands(1379) <- agent_comms.cli(1594)
agent_comms.tool_declaration(6) <- agent_comms.native_tools(59) <- agent_comms.threads(283) <- agent_comms.input_disposition(294) <- agent_comms.goal_actions(508) <- agent_comms.registration(513) <- agent_comms.historical_views(523) <- agent_comms.message_bus(736) <- agent_comms.cli_commands(1379) <- agent_comms.cli(1594)
agent_comms.agent_events(37) <- agent_comms.pi_payloads(77) <- agent_comms.native_entries(115) <- agent_comms.native_transcript(129) <- agent_comms.external_transcript(135) <- agent_comms.unstarted_runtime(136) <- agent_comms.session_file(140) <- agent_comms.threads(283) <- agent_comms.input_disposition(294) <- agent_comms.goal_actions(508) <- agent_comms.registration(513) <- agent_comms.historical_views(523) <- agent_comms.message_bus(736) <- agent_comms.cli_commands(1379) <- agent_comms.cli(1594)

agent_comms.commander.cli

acp(254) <- acp.schema(254) <- agent_comms.controller(254) <- agent_comms.input_origin(257) <- agent_comms.input_attempt(262) <- agent_comms.acp_failure(333) <- agent_comms.agent_events(353) <- agent_comms.pi_events(538) <- agent_comms.plan_phases(573) <- agent_comms.heartbeat_frame(614) <- agent_comms.commander_heartbeat(812) <- agent_comms.commander.standing(1129) <- agent_comms.commander.order(1475) <- agent_comms.commander.cli(1681)
agent_comms.tool_declaration(5) <- agent_comms.pi_payloads(58) <- agent_comms.acp_failure(333) <- agent_comms.agent_events(353) <- agent_comms.pi_events(538) <- agent_comms.plan_phases(573) <- agent_comms.heartbeat_frame(614) <- agent_comms.commander_heartbeat(812) <- agent_comms.commander.standing(1129) <- agent_comms.commander.order(1475) <- agent_comms.commander.cli(1681)
agent_comms.agent_events(353) <- agent_comms.pi_events(538) <- agent_comms.plan_phases(573) <- agent_comms.heartbeat_frame(614) <- agent_comms.commander_heartbeat(812) <- agent_comms.commander.standing(1129) <- agent_comms.commander.order(1475) <- agent_comms.commander.cli(1681)

agent_comms.cli_bridges.claude

agent_comms.tool_declaration(6) <- agent_comms.native_tools(51) <- agent_comms.threads(541) <- agent_comms.private_send_rules(544) <- agent_comms.native_input_owner(585) <- agent_comms.agent_activity(589) <- agent_comms.runtime_updates(616) <- agent_comms.cli_bridges.claude(1370)
acp(253) <- acp.schema(253) <- agent_comms.controller(253) <- agent_comms.input_origin(256) <- agent_comms.input_attempt(261) <- agent_comms.acp_failure(277) <- agent_comms.agent_events(298) <- agent_comms.pi_payloads(333) <- agent_comms.native_entries(375) <- agent_comms.native_transcript(389) <- agent_comms.external_transcript(395) <- agent_comms.unstarted_runtime(396) <- agent_comms.session_file(399) <- agent_comms.threads(541) <- agent_comms.private_send_rules(544) <- agent_comms.native_input_owner(585) <- agent_comms.agent_activity(589) <- agent_comms.runtime_updates(616) <- agent_comms.cli_bridges.claude(1370)
agent_comms.agent_events(298) <- agent_comms.pi_payloads(333) <- agent_comms.native_entries(375) <- agent_comms.native_transcript(389) <- agent_comms.external_transcript(395) <- agent_comms.unstarted_runtime(396) <- agent_comms.session_file(399) <- agent_comms.threads(541) <- agent_comms.private_send_rules(544) <- agent_comms.native_input_owner(585) <- agent_comms.agent_activity(589) <- agent_comms.runtime_updates(616) <- agent_comms.cli_bridges.claude(1370)

The direct schema edges from tool_declaration and agent_events are gone; the remaining commander/bridge chain still passes transitively through agent_events -> acp_failure -> input_attempt -> input_origin -> controller. The controller/input_origin dependency was explicitly left for the next cut.

Interpreter and static checks

All seven declarations resolve to the original SDK answers: ReadTool->ReadAction->read; BashTool->ExecuteAction->execute; EditTool/WriteTool/FilePatchTool->EditAction->edit; GrepTool/GlobTool->SearchAction->search. The registry contains five capability declarations, not tool records. The unclassified base resolves to OtherAction.

ImageInput('YQ==', 'image/png') retains {'data': 'YQ==', 'mimeType': 'image/png', 'type': 'image'} and decodes equal. The paired Toad prompt, clipboard and resource modules import; saved ReadAction publishes as read.

Mypy --explicit-package-bases --follow-imports=silent on the six existing declaration/projection modules: 77 -> 73 errors, no file grows. New image_input.py was also checked without errors. Ruff on all 19 touched Core files: 87 -> 67 findings, no file grows. ToolAction is intentionally an ABC marker per the owner decision; it has no fake abstract behavior merely to silence B024. No pytest or other test runs.

Checked module	Mypy before	Mypy after
src/agent_comms/agent_event_updates.py	1	1
src/agent_comms/agent_events.py	8	8
src/agent_comms/image_inputs.py	0	0
src/agent_comms/native_tools.py	19	15
src/agent_comms/tool_declaration.py	5	5
src/agent_comms/transcript_events.py	44	44

Assumed:

• Core Keep ACP vocabulary out of native tool and image declarations #2913 lands before Toad P167: let real input changes own drain recovery #1165; paired installation uses both updated heads.
• Direct ToolAction subclasses declare action capabilities; tool subclasses inherit them and do not become new action categories.
• OtherAction remains the answer for a tool with no declared action capability, as ordered.

MCP field collision correction

Head 10b3ebb merges origin/main 6b042c7 and renames the capability classmethod to declared_action without changing its body. This avoids CommsCollaborationTool.action, a required MCP field whose name must stay unchanged. All three callers are migrated. The only remaining .action() call is coordinated_runtime.py:132; the declared_action field/assignment grep has no matches. No test runs.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.