fix(computer-use): stop dropping tool name when action is present

[gzak]

Summary

_derive_connector_type(tool_name, action) in axonflow/adapters/computer_use.py discarded tool_name whenever action was present, returning only f"computer_use.{action}". Every actioned Computer Use tool call (e.g. computer with action="left_click") collapsed to the same connector_type shape regardless of which actual tool triggered it — connector_type couldn't distinguish tool identity for any actioned tool.

Fixes #2910 (epic #2905). Uses the platform's two-field (server, tool) identity contract (platform sub-issue #2904, merged to axonflow-enterprise as c8df2006b and first released in platform v9.10.0).

Design decision (epic #2905, RULING 1)

Per RULING 1 on the epic, tool carries the tool name in every adapter so the tool_name audit column means the same thing regardless of which adapter emitted it. There were three pieces of information: the "computer_use" domain marker, tool_name ("computer"/"bash"/"text_editor"), and action ("left_click"/"screenshot", present only for some tools). They map as:

• connector_type = the constant "computer_use" — the connector/domain marker, analogous to the LangGraph adapter's connector_type = server_name. Restoring "computer_use" as the connector value also keeps computer_use-scoped policies matchable on connector_type.
• tool = tool_name (e.g. "computer") — the tool NAME, never the action. This is what makes the identity coherent with the LangGraph adapters (where tool is also the tool name).
• action stays in statement — json.dumps(tool_input) already serializes the action key, and the policy engine evaluates the statement. The action is not placed in tool, and not in operation: operation is a spec-declared enum {query, execute} in agent-api.yaml, so putting "left_click" there would violate our own published contract. operation stays "execute".

_derive_connector_type is removed. check_result() sends the same identity on the response plane (connector_type="computer_use", tool=tool_name) so request- and response-plane rows correlate — the platform doesn't yet consume tool on check-output (#2955), so it is sent forward-compatibly.

This is a breaking change (the connector_type wire value changes shape) and must ship as a major SDK bump; this adapter has no connector_type_fn escape hatch, so re-scoping policies is the only migration path — see the CHANGELOG.

Also (shared with #2906, byte-identical hunks — merge #216 first):

• mcp_check_input/mcp_check_output (and the check_tool_input/check_tool_output aliases), async and sync, gain an optional tool: str | None = None parameter, sent on the wire only when non-empty.
• MCPCheckInputRequest / MCPCheckOutputRequest gain a matching optional tool field.
• tests/fixtures/wire_shape_baseline.json acknowledges tool as an SDK-superset field on both request models (the public agent-api.yaml spec hasn't declared it yet, though the agent consumes it on check-input since v9.10.0).

Test plan

• Rewrote TestConnectorTypeToolSplit to assert the RULING-1 mapping: connector_type == "computer_use" (constant) and tool == the tool name for actioned and non-actioned tools; that two calls to the same tool with different actions carry the same (connector_type, tool) identity while the differing action lives in statement; and that operation stays "execute" (spec-compliant, action never smuggled in).
• Updated TestCheckToolUse / TestCheckResult assertions to the new mapping (connector_type="computer_use", tool=tool name on both planes).
• Rewrote runtime-e2e/computer_use_tool_name_split/test.py for the new mapping (wire-body capture against a live agent; action asserted present in statement).
• pytest tests/test_computer_use.py — 30 passed; full suite — 1052 passed, 29 skipped.
• ruff check / ruff format --check on changed files — clean; mypy axonflow — no new errors.
• Regenerated .lint_baselines/falsey_clobber.json (line-shift only). Merged origin/main (the caller_name legs) in non-destructively; CI green, MERGEABLE/CLEAN.

[saurabhjain1592]

R3 review (master + independent hostile pass). The core fix is verified correct — previously _derive_connector_type discarded the tool name whenever an action existed; now both survive, with good unit + live-wire coverage (actioned/non-actioned/same-tool-different-action/check_result). Must-fix:

BLOCKER — CI is red

lint fails (run 29540706653) with 12 NEW falsey-clobber findings, mostly in files this PR doesn't touch (client.py:1682/1723/1781/2299, computer_use.py:175/209) — line-number baseline drift. Regenerate .lint_baselines/falsey_clobber.json in this PR.

MEDIUM — semantic drift needs an epic-level ruling before merge

This PR sends tool = ACTION (left_click) with connector_type = tool name (computer, bash), while the four langgraph PRs send tool = tool name. On a new platform, audit rows here read tool_server="bash" / tool_name="left_click" — bash is not a server, and the meaning of the tool_name audit column now depends on which adapter emitted it. If intended, record the decision on epic #2905; if not, tool should carry the tool name and the action belongs elsewhere (e.g. operation/statement).

MEDIUM — stale contract claims + #216 collision

wire_shape_baseline.json note + types.py comment say #2904 is "not yet merged" — it merged 2026-07-16 (enterprise c8df2006b). Update, and reconcile with #216's edits to the same lines (guaranteed conflict; #216's text is the correct one).

MEDIUM — hard policy break with no escape hatch

connector_type loses the computer_use. namespace entirely (computer_use.bash → bash) on EVERY platform version, so existing policies/per-connector settings scoped to computer_use.* stop applying — and this adapter has no connector_type_fn escape hatch. The changelog migration note is the only mitigation; it's currently absent. Add it (Go #189 has the model text).

[saurabhjain1592]

Delta note — semver (net-new, strongest case in the cluster). This drops the computer_use. namespace from connector_type entirely (computer_use.bash -> bash) on every platform version, so all existing computer_use.*-scoped policies and per-connector settings silently stop applying — and this adapter has no connector_type_fn escape hatch to restore the old value. That is a textbook incompatible behavioral change with no migration lever: MAJOR bump required, or add an escape hatch. Current changelog (### Fixed, no migration note) materially understates it.

Also still open at head: CI is RED (falsey-clobber baseline line-drift, 12 new + 12 stale — baseline never regenerated in-PR); the tool=ACTION vs tool-name semantic still needs the epic ruling (item 3 on the #2905 roll-up — no ruling posted yet); the "#2904 not yet merged" comments are now stale (#2916 shipped in v9.10.0).

[saurabhjain1592]

Follow-up commits (WS-DECONCAT): applied epic #2905 RULING 1 — connector_type="computer_use" (constant), tool=the tool NAME (computer/bash/text_editor, never the action). The action is NOT put in tool and NOT in operation (agent-api constrains operation to the enum {query,execute}) — it stays in the JSON statement, which already carried it. check_result sends the same identity. _derive_connector_type removed. Reconciled the shared client/types/baseline hunks with #216 (byte-identical) and regenerated the falsey_clobber baseline. Changelog is Changed (BREAKING) (major bump, no connector_type_fn escape hatch on this adapter — called out, computer_use.* namespace-loss migration note, min-platform pins). Merge #216 first. Merged main in. CI green, MERGEABLE/CLEAN, master-R3'd. Ruling recorded on #2905.

[saurabhjain1592]

Master independent R3: CLEAR-FOR-SET (code). RULING 1 verified correct: connector_type="computer_use", tool=tool name, action stays in statement (pinned by test_tool_name_survives_regardless_of_action). The operation-enum reasoning is TRUE — agent-api.yaml operation is enum:[query,execute], so the action can't go there; adapter sets operation="execute". Namespace-loss/no-escape-hatch BREAKING changelog prominent; shared hunks byte-identical with #216; CI green; non-destructive.

MEDIUM (non-code) — update this PR's description. The body still describes the PRE-RULING-1 mapping (connector_type=tool_name, tool=action, "different actions → different tool values") — the exact inversion the shipped code overturned. Code/tests/CHANGELOG are all correct; only the description is stale and would mislead a reviewer trusting it. Fix before merge.

[saurabhjain1592]

Release-ready follow-up: bumped the version to 9.0.0 and finalized the tool-identity content under a real ## [9.0.0] - 2026-07-18 changelog header (dropped the "must be a major bump" banner — the header is the bump). Scrubbed the [9.0.0] changelog of internal issue/PR numbers and internal review terminology; minimum platform is now stated as versions (check-input tool = AxonFlow platform v9.10.0+, response-plane tool = v9.11.0+). Version-alignment gate green. CI green, MERGEABLE/CLEAN. Master-R3'd (2 rounds, incl. an independent finalization pass).

Python: #216 and #217 share one byte-identical, combined [9.0.0] de-concat entry — merge #216 first. (PR description also rewritten to the shipped mapping.)

[saurabhjain1592]

Master re-R3 of the v9.0.0 finalize delta: CLEAR-FOR-SET. Version file 8.5.1 → 9.0.0; correct Keep-a-Changelog shape (empty [Unreleased] above ## [9.0.0], BREAKING content under [9.0.0], release-workflow preflight will match). Public [9.0.0] block scrubbed clean — no enterprise issue/PR numbers, no RULING language, min-platform stated as versions (v9.10.0+/v9.11.0+). Finalize commit is additive; de-concat logic (two-field split + rulings) untouched. CI green.