[jothimani-rajendran]

What this changes

This is batch A3 of org-plan plan/one-oh-gaps.md (at 5271b89). It adds four matrix rows, all at basis: vendor-docs.

• openhands gets a real, config-only adapter in the flat_decision family.
  • src/agentseam/data/vendors/openhands.json is added and joins _CONFIG_DRIVEN.
  • A golden fixture, tests/test_adapter_openhands.py (14 tests) and examples/generated/openhands.md come with it.
  • No core or engine change was needed.
• amp, kiro and factory_droid get unadapted rows with events {}, dated doc links and a note saying why.

Derived tiers at pre_tool, from agentseam tier-table. The table's own Tier column is the matrix's categorical tier field, which is set in the row; the grade and fail mode are derived.

Agent	Tier	Grade	Fail mode	Basis	Recorded
openhands	block	best-effort	open	vendor-docs	2026-10-11
amp	unadapted	none	n/a	vendor-docs	2026-10-11
kiro	unadapted	none	n/a	vendor-docs	2026-10-11
factory_droid	unadapted	none	n/a	vendor-docs	2026-10-11

openhands' grades at its other events: best-effort at prompt_submit and stop; detect at post_tool, session_start and session_end.

Evidence-wanted issues, each linked from its row's notes:

• Evidence wanted: openhands adapter (vendor-docs) #181 (openhands)
• Evidence wanted: amp adapter (vendor-docs) #182 (amp)
• Evidence wanted: kiro adapter (vendor-docs) #183 (kiro)
• Evidence wanted: factory_droid adapter (vendor-docs) #184 (factory_droid)

Claim check

• No capability claim is widened without a mechanism behind it
• Every new MATRIX row carries a verified record (version, date, method)
• Payload shapes come from a primary source:
  • OpenHands: https://docs.openhands.dev/openhands/usage/customization/hooks, https://docs.openhands.dev/sdk/guides/hooks, and the SDK's file_editor/definition.py at v1.54.0. Versions: OpenHands v1.26.0 (2026-10-08); openhands-sdk 1.54.0, measured on PyPI (2026-10-09).
  • Amp: https://ampcode.com/docs/customize/plugins, https://ampcode.com/notes/permissions, https://ampcode.com/docs/cli/settings and https://ampcode.com/news/neo. Version: @ampcode/cli 0.0.1791676897-g77c125, measured on npm (2026-10-11).
  • Kiro: https://kiro.dev/docs/hooks/, https://kiro.dev/docs/hooks/types/, https://kiro.dev/docs/hooks/actions.md and https://kiro.dev/docs/cli/v3/hooks-migration.md. Versions: IDE 1.2.56 (2026-10-08) and CLI 2.29.0 (2026-10-09), from the changelogs.
  • Factory Droid: https://docs.factory.com/harness/hooks and https://docs.factory.com/harness/agents-md. Version: droid 0.238.0, measured on npm (2026-10-10).

Why each of the three is unadapted

• Amp. The only hook surface is an in-process TypeScript plugin API, which OPERATING keeps on watch. The legacy delegate permission helper is an exit-code check. It receives no event name and runs only when the user has already opted into the old permissions system.
• Kiro. It has command hooks, but in a dialect no family speaks:
  • the config is {version: "v1", hooks: [{name, trigger, matcher, action: {type, command}}]}, which no install wrapper writes;
  • the verdict is an exit code only, and the IDE and CLI read it differently;
  • the payload has no key that separates it from Claude Code's.
• Factory Droid. It speaks the Claude Code dialect, but its payload carries Claude Code's own keys, including transcript_path and permission_mode, which looks_like_claude_code treats as markers. A factory_droid claim would make detection ambiguous for every Claude Code payload too, and an ambiguous payload is allowed through. Separating the two needs a value-based probe, which is a core change.

Checks

• pytest -q passes locally: 2117 passed, 4 skipped
• ruff check . and ruff format --check . pass
• The runtime path is still stdlib-only (tests/check_stdlib_only.py OK)
• Commits are signed off

Notes for the reviewer

Measured risk, recorded in the kiro row rather than fixed (it would need a core change). Today detect() assigns a Kiro-shaped payload to claude_code. A deny then goes back as a JSON body with exit 0, and Kiro reads only the exit code, so the deny would not block. The row says Kiro is not a supported target. A Droid-shaped payload is also detected as claude_code, and its wire answer is the one Droid documents. However, Droid's Execute and Create tools are not in Claude Code's tool lists, so a policy keyed on shell or write tools would not recognise those calls. The factory_droid row records this.

Known OpenHands limit, recorded in the row. file_editor carries its own sub-command (create, str_replace and so on) in tool_input.command, the same key the shell tool uses. So Event.command carries it for file-editor calls.

Changes beyond the plan's Touches, all required by invariant tests or CI generators:

• Per-agent data: instructions.py (AGENTS.md for all four agents), allow-semantics for openhands, and the packaging and permissions unrecorded reasons.
• Capture footprint: an openhands entry in tools/capture.py.
• Recount entries: in tables.py and dialects.py.
• Test probe: the test_decision_vocabulary probe adds event_type=.
• Pins: the test_instructions pin moves from 21 to 25, as do the make_matrix and make_social pins.
• Regenerated: the figures and the social preview.
• Re-measured counts:
  • wiring counts in the event_log.py and notify.py docstrings: 11 of 14 and 13 of 14 adapters wire;
  • the adapter counts in tools/VERIFY.md and tools/capture.py: 12 of 14 adapters rest on docs or source.

README counts are left alone on purpose: A4 owns them.

Missing label: evidence wanted does not exist in this repo. The four issues carry help wanted only. Should the owner create the label?

🤖 Generated with Claude Code