ADFA-6311 | Add AI-Agent-Claude, a Claude backend for AI Core

[davidschachterADFA]

ADFA-6311

Adds plugins/AI-Agent-Claude: a claude backend for AI Core over Anthropic's Messages API, with streaming, history, native tool calling, Stop, and a settings pane (Keystore-encrypted key, model picker from /v1/models, connection test).

Review by commit

1. 52101c4 — mechanical: AI-Agent-OpenAI copied and renamed (package, classes, resource prefixes, Keystore alias). Skim it.
2. 8c0a443 — the change: Messages API transport and the trimmed settings pane. Diff it against commit 1 to see exactly what Claude needs.
3. 4a37a15 — docs: README, Tier 3 guide, overview page, root README row, skip.txt.
4. 665bb0e — icons from the Claude symbol (Wikimedia Commons, CC0; the mark is Anthropic's trademark).
5. 4b3a48c — found on device: a key with no workspace gets a 400 on every request; the key check called that "Unknown" and blamed a missing AI Core. Now its own failure and verdict.
6. 383fe1d — an optional Workspace ID field, shown only after Claude refuses a key for that reason; sent as anthropic-workspace-id, saved and cleared with the key. Typed IDs are validated so a pasted value cannot add a header.
7. b4b2703 — every prompt for the ID says it starts with wrkspc_ and is listed in the Claude Console under Workspaces.
8. da01fac — fixes from an independent review (below).
9. 3463594 — the docs that review flagged.
10. 0266fc3 — the gallery card (addon.json). The addon stays held in tools/addons/skip.txt, like AI-Code-Suggestions, until AI Core is in the gallery.

Review findings and how they were handled

Finding	Handling
F06/F07 per-model fields	Each model's output cap, adaptive thinking and effort now come from /v1/models, stored with the model list; allow-lists cover undescribed models. Agent turns ask for adaptive thinking where it's taken (on Opus 4.x / Sonnet 4.6 an omitted thinking meant none). Dated Opus 4 / Sonnet 4 ids no longer get effort or an over-cap budget.
F08 small jobs	Titles and suggestions keep the caller's budget at low effort; only the 5.x models get a 4K floor.
F09 aliases	An alias matches its dated id, so a live fetch no longer swaps claude-haiku-4-5 for Opus 5.5.
F10/F15	A fallback drops the declined model's tool calls; a call cut off by max_tokens is dropped.
F11/F12 Stop	Stop closes the stream's socket at once and tracks every stream.
F13	Test Connection adopts the catalog of the key it tested.
F14	A silent stream is reported as Claude stopping, not a lost connection.
F16 tests	TurnAssembler (pure) decides how a turn ends; it and the retry loop are tested.
F17/F18	Docs and comments corrected.
F19	Not changed: the copy commit registers as openai, which only matters when bisecting; pushed history left alone.
F20	Decided 2026-10-02: proceed treating use of the Claude name and symbol as permitted. Anthropic's Trademark Guidelines require prior written approval and their own unaltered asset, so confirm the permission is actually in hand before the gallery listing, and swap in the supplied asset if it comes with one.

Design choices

• Raw HttpURLConnection, no Anthropic SDK. The host loads plugins parent-first, so a bundled OkHttp collides with the host's (same reason as the OpenAI backend).
• History is text. AI Core hands tool results back as user text, never the tool_use they answer, so they are sent as text, not tool_result blocks. Same-role turns are merged; no thinking block is ever replayed.
• No thinking, no temperature. Both are 400s on Claude Opus 5.5, the default model. max_tokens is raised to 64K for streams because thinking counts against it.
• Per-model fields (ClaudeModelTraits): effort: "high" where accepted; server-side fallbacks: "default" on the models whose safety classifiers can decline a request.
• stop_reason: "refusal" is reported before any tool call in that turn runs.
• Retries: 429/529/5xx retried twice, only before a stream has shown anything.
• No embeddings: Anthropic has none, so this is not an EmbeddingBackend; Vector-Search still needs OpenAI or Gemini.
• Not strict tools, no eager input streaming: contributed MCP schemas rarely meet strict mode, and calls are acted on only once complete, so eager streaming buys nothing and loses server-side input validation.

Verification

• 161 JVM tests pass; assemblePlugin builds. Mutation checks: removing turn merging, the Haiku effort gate, the credit-balance mapping, the workspace detection, the header-safety check, alias matching, the Opus 4 cap, the fallback discard, the max_tokens drop, adaptive thinking on agent turns, low effort on small jobs, or the stall classification each fails the tests named for it.
• On device after the review fixes (build 3463594): the live catalog's capabilities are stored and match the fallback tables; a saved claude-haiku-4-5 survives Test Connection; chat plus title request accepted on Haiku 4.5, Opus 4.8 (adaptive thinking + high effort, title at low effort) and Opus 5.5; an Opus 5.5 agent edit ran read_file → edit_file → respond; Stop closed the stream's HTTPS socket within 2 s (checked in /proc/net/tcp6).
• On device (Galaxy Note20 Ultra, Android 13, CoGo debug 2026-10-01, with AI Core from main), against the live API:
  • Install through CoGo's installer, upgrade via "Replace", restart; backend registers. Claude icon in the Plugin Manager, day and night.
  • Invalid key: real 401, "Claude refused this key", nothing saved.
  • Key without a workspace: refused with the workspace message, field appears; with the ID, key verified. Test Connection lists 12 models.
  • Chat on claude-opus-5-5: streamed reply, end_turn. Agent edit of MainActivity.kt via search_project → read_file → edit_file → respond; the comment is on disk.
  • Stop mid-stream: turn cancelled, no error, no late completion.
  • claude-haiku-4-5: accepted without effort/fallbacks, served by claude-haiku-4-5-20251001.
  • Empty credit balance: shown as the billing message.
  • Offline save: "Save anyway" stores the key unverified; re-saved verified once online.
  • Tooltips on the key, model, Get API Key and Test Connection: Claude wording, never "n/a".
  • Font scale 1.0 and 2.0: settings pane (view and edit mode, with the workspace field) wraps and scrolls; every control reachable. Only the field's sample hint is ellipsized.

🤖 Generated with Claude Code

[davidschachterADFA]

CI build passed for AI-Agent-Claude at 3463594: Build plugin artifacts, run 37076228947, dispatched with plugin=AI-Agent-Claude and codeonthego_ref=stage.

• It rebuilt plugin-api from current CoGo stage instead of using the committed libs/ jars, and produced a release .cgp.
• It does not run unit tests. The 161 JVM tests have only run locally (../../gradlew testDebugUnitTest).

🤖 Generated with Claude Code