Repository navigation
Conversation
Review every bullet of the SDK map (S01..S21) and the agent-server map (F01..F34) and link the ones that prove the same promise on both surfaces: 136 SDK bullets now end with `Server:` naming 229 agent-server bullets, which end with `Same promise:` naming them back. Define the link contract in both map indexes, make `sdk-map check` and `control-agent-server map check` validate the links in both directions (target exists, names the bullet back, exact form), add `sdk-map links` and `control-agent-server map links` to list the pairs, and add a twins step to both daily passes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013ArwwJcJzCzwmKD5zoVnd1
REST API breakage checks (OpenAPI) — ✅ PASSEDResult: ✅ PASSED |
all-hands-bot
left a comment
There was a problem hiding this comment.
This review was posted by an AI agent (OpenHands).
Reviewed at head 3a97f03e89591dd4c9bcc92797de8857a78fa060 (55 files, +840/−369). Scope gate: the change touches only .agents/skills/verify-sdk/, .agents/skills/verify-agent-server/, and tests/cross/, all owned by this repository; linked issue #6727 is ready-for-dev. No public API, REST contract, settings, events, or persisted-state surfaces are touched, so the compatibility checkpoints do not apply.
What I verified on this head:
- Both tools validate the twin links bidirectionally and pass:
sdk-map check→ ok, 0 problems (458 IDs);control-agent-server map check→ ok, 0 problems (953 IDs, routes included). Link counts match the PR description exactly: 136 SDK bullets withServer:links naming 229 agent-server bullets, 268 pairs in each direction; zero malformed links, unknown targets, duplicates, or wrong-prefix IDs. - Spot-read a dozen pairs in full text (S04.pause ↔ F05./F33, S13.confirm- ↔ F08/F33, S19.profile-store ↔ F21., S03.meta-profile-store ↔ F22., S02.error-mapping ↔ F20.validate-, S17.usability ↔ F27.catalog, S10.command-exec ↔ F27.hooks-conversation, S20.conversation-args ↔ F04.): each is genuinely the same promise on both surfaces, matching the documented contract. The claimed removals (S14.terminal-exec, S13.llm-analyzer, S19.acp-settings) indeed have no links, and the claimed additions (S02↔F20, S04↔F33, S08↔F25) exist.
- The S01.stuck-recover ↔ F05.stuck-rerun pair links a bullet that self-declares a known failure (#6648); this is consistent with the daily-pass rule that a disagreeing pair is a finding "unless the difference is already a known bug on one side", so it is not a defect.
- Error paths behave as the tests assert: malformed link form, unknown target, one-sided link,
--filefiltering, and a checkout without the sibling map (skips cleanly) — all exercised directly, not just via the tests. - Diff integrity: every removed map bullet line reappears with only a link appended (0 mismatches across 365 replaced lines); nothing else in the maps was rewritten. Ruff check/format clean on the changed scripts; doc anchors and the new daily-pass commands parse against the real CLIs.
- Tests:
uv run pytest tests/cross/test_verify_sdk_skill.py tests/cross/test_verify_agent_server_skill.py→ 97 passed, matching the PR claim. - Checks on this head are green; the one failing "Validate PR description" run is stale — the same check has later successful runs on this exact SHA, and the REST API breakage bot reports PASSED at 2026-11-10T17:42Z.
No material correctness, security, or compatibility findings.
✅ APPROVED
Resolve conflicts by taking main's map text and re-applying the PR's Server:/Same promise: links (all 365 linked bullets still exist on main). Link 28 new twin pairs main added (model router, ask_user, message edits, reactions, runtime MCP add, ACP effort, public-skills switch, ...), and move F09.load-plugin-twice to main's new S10.plugin-hooks-reload-once. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013ArwwMeAwT1A2JVHapmA1X
all-hands-bot
left a comment
There was a problem hiding this comment.
"This review was posted by an AI agent (OpenHands).\n\nFresh review of head bd418bc52b6492dd6ce4f08d9da30866dd68a713 (the prior approval from this account was on 3a97f03e, before the merge of main).\n\nScope: entirely within .agents/skills/verify-sdk/, .agents/skills/verify-agent-server/, and tests/cross/ \u2014 SDK/agent-server verification tooling and maps, owned by this repository. No production SDK/server code paths, no API or compatibility surfaces, no secrets.\n\nVerified on the current head in a clean checkout:\n- sdk-map check \u2192 ok, 0 problems (21 files, 504 ids); control-agent-server map check \u2192 ok, 0 problems (38 files, 1068 ids), routes included.\n- sdk-map links --id S04.pause \u2192 exactly the 6 pairs the PR claims; control-agent-server map links --file F08 works as documented.\n- Both tools agree on the full pair list (295 edges, each direction naming the other back).\n- uv run pytest tests/cross/test_verify_sdk_skill.py tests/cross/test_verify_agent_server_skill.py \u2192 100 passed (the PR body cites 97; the count grew because the head merged newer main \u2014 all pass).\n- All GitHub Actions checks on this head are green (cross-tests, agent-server suites, OpenAPI, builds).\n\nLink contract: the bidirectional check is strict \u2014 the named ID must exist in the sibling map, must name the bullet back, and the link must end the line in the exact documented form; duplicate IDs, unknown targets, and malformed lines are all reported (exercised directly, including through the new tests covering one-sided links, unknown targets, malformed links, the --file filter, and the missing-sibling skip). Spot-checked claimed pairs (S02.error-mapping \u2194 F20.validate-*, S08.marketplace-auto-load \u2194 F25.catalog-registered-marketplace, S04.pause \u2194 F05.*/F33.conversation-pause): the linked bullets do state the same promise on both surfaces.\n\nFreshness: the branch merged main at 3d638f2e (including map maintenance #7006) and the merge commit linked the new twins; the three main commits since (ruff bump, release, ACP fix) do not touch any linked bullet lines, and the checks above run against current main content. The PR body's 136/229 counts predate that merge (the merged head carries more linked bullets), which is description staleness only \u2014 the map data itself validates consistent. The git status note about the stranded cloud/sweet-darwin-b4uwcp branch still applies and is correctly flagged for a rebase by whichever lands second.\n\nIssue #6727 remains open with ready-for-dev/priority:low; this PR delivers its acceptance criteria as described.\n\nNo material correctness, security, compatibility, or acceptance-criterion defects found.\n\n\u2705 APPROVED\n"
HUMAN:
start a new thread to solve this, go through everything in sdk and server maps, and link them appropriately "There were no SDK/server twin pairs to drive, because the map has no Server: links yet."
AGENT:
Requested by Engel · project thread
I'm an AI agent (Claude Code) helping Engel Nyst (@enyst) with project maintenance.
Why
The daily map routine drives SDK/agent-server twin pairs (an SDK bullet and an agent-server bullet that prove the same promise), because twins that disagree usually locate a bug at the seam between the SDK and the server. No bullet carried a link, so the 2026-11-10 run drove no pairs.
Summary
Server: ..., naming 229 agent-server bullets, which end withSame promise: ...naming them back. A link means "a pass driving both expects the same answer", not "related feature": server-only contract (status codes, auth, stores, webhooks) and SDK-only behavior (local workspaces, visualizers, logging, LLM internals, critic, planning tools) stay unlinked. The 2026-11-03 overlap table was the starting point; I re-checked each pair against current main, tightened loose ones (e.g.S14.terminal-exec,S13.llm-analyzer,S19.acp-settingshad no true twin) and added missing ones (e.g.S02.error-mapping↔F20.validate-*,S04.pause↔F33.conversation-pause,S08.marketplace-auto-load↔F25.catalog-registered-marketplace).mapping.md.sdk-map checkandcontrol-agent-server map checkvalidate links both ways: the named ID exists in the other map, names this bullet back, and the link ends the line in the exact form. Newsdk-map links/control-agent-server map links [--file] [--id]list the pairs.daily.mdfiles get a step 3, Twins: drive the twins of the families a change touched, and treat a disagreeing pair as a finding.Issue Number
Fixes #6727
How to Test
New tests cover a one-sided link, an unknown target, a malformed link, the
--filefilter, and a checkout without the sibling map (checks skip).Type
Notes
Surfaces: sdk, server(running an SDK bullet throughcontrol-sdk --surface remote) is not part of this PR, because that runner doesn't exist yet. The 253 SDK bullets that run inside the server but have no driving F bullet stay unlinked until it does.cloud/sweet-darwin-b4uwcp) touch a few of the same bullet lines (S04, F07, F15); whichever lands second needs a small rebase.🤖 Generated with Claude Code
https://claude.ai/code/session_013ArwwJcJzCzwmKD5zoVnd1
🐳 Agent Server images for this PR — GHCR package, pull/run commands, and all pushed tags (click to expand)
• GHCR package: https://github.com/OpenHands/agent-sdk/pkgs/container/agent-server
Variants & Base Images
eclipse-temurin:17-jdkpython-node-runtimepython-node-runtimepython-node-runtimegolang:1.21-bookwormPull (multi-arch manifest)
# Each variant is a multi-arch manifest supporting both amd64 and arm64 docker pull ghcr.io/openhands/agent-server:bd418bc-pythonRun
All tags pushed for this build
About Multi-Architecture Support
bd418bc-python) is a multi-arch manifest supporting both amd64 and arm64bd418bc-python-amd64) are also available if needed