ta
TanStack
OSS
ta
Pipelines
Nx runs
Insights
Compare tasks
Analytics
Sign in
Toggle Sidebar
ai
ai
Pipelines
fix(ai, ai-harness): a guard tier for tool calls, complete plugin removal, media and body limits, signing keys, and turn and config fixes (#1824) * fix(ai-harness): a settled operation keeps its status, a media signer with key rotation, and agents() rule docs that match the code * fix(ai-harness): every plugin contribution has an owner, ctx.on cleans up, inline media goes to the media store, and capabilities report the socket only when mounted * docs(ai-harness): say when capabilities report the WebSocket * feat(ai): onGuardToolCall middleware hook that runs on the final arguments of every tool call * test(e2e): onGuardToolCall denies on the final args after a transformArgs * fix(ai-harness): tool turns take no steers, a turn keeps its config, removing a plugin removes all of it, body and frame limits, signing keys, and a warning on tool name clashes * fix(ai-harness): the media signer's key check covers an empty secret * fix(ai-harness-cli): serve signs media URLs with a key derived from the token, so a short token meets the 32-byte key rule * feat(ai-harness, ai-mcp, ai-acp, ai-code-mode, ai-harness-cli): the final review of the harness, with its fixes, one object per call, and the known limits (#1825) * fix(ai-harness, ai-mcp, ai-acp, ai-code-mode): final review fixes for durability, access, plugins, transports, integrations, and coding tools * fix(ai-acp): one agent app serves one connection, with editor capabilities read from that connection, and no private SDK field * fix(ai-harness): a thread delete ends it in first-party plugins, the bash spill is scrubbed, removed schedules go at once, agent run receipts resolve, and session.append takes one object with an optional inputId * refactor(ai-harness, ai-mcp, ai-harness-cli): session, client, and view methods take one object, the client and view use putMedia and client.sessions, and internal fields leave the public types * docs(ai-harness, ai-mcp, ai-acp, ai-code-mode): the final review changes, the fixed durable-tools and server-tools text, and a known limits section * fix(ai-harness): revoke and unshare reject when work does not stop, thread delete reaches child logs and known harnesses, plugins declare kept records, and the review fixes - host.revoke and host.sessions.unshare reject when the work stop fails (the access is gone already). unshare ends the reads before it stops the work. - createHarnessHost takes harnesses, so a delete on a new host ends the thread in its plugins. An unknown harness warns. A child thread ends in the log of its root. - A plugin declares keepRecords. The core fold reads them from the plugins of the harness that folds the log, and counts per type. snapshots() keeps its 500 steps this way. - revert and unrevert open the workspace of the thread from the log when no turn opened it on this host. - The harness.revert and harness.job records keep the principal. - append with no inputId goes through accept(), so its id is real. - A formatter config below the root, or a write under node_modules, asks again. - The base64 cache takes the largest encodedBytes. The client takes idleTimeoutMs. - An events() read listens for a revoke from the call, not the first pull. * docs(ai-harness): the harnesses option, keepRecords, idleTimeoutMs, a revoke that rejects, and one known limit less * feat(ai-harness): ctx.session.keptRecords reads the kept records of a plugin, and session.revert reads them the same way * fix(ai-harness): an always answer saves no rule for an @unsaved: action * refactor(ai-harness): mounted plugins keep each contribution list with its owner, not in parallel owner arrays * fix(ai-harness): a plugins remove setting is rejected when a kept turn plugin or a harness middleware needs the removed capability * fix(ai-harness): client media checks every part before it stores any, refuses provider file handles unless media.allowProviderFiles, and the frame limit docs name the server maxPayload * fix(ai-harness): a retried client input with inline media reuses the stored media, so it stores nothing again and is a duplicate, not a conflict * docs(ai-harness): routed root-agent turns take steers and late joins, only session.tool() turns do not * fix(ai-harness-cli): --serve signs media URLs with --media-key or HARNESS_MEDIA_KEY, or a random key, never with the bearer token
nx run-many --targets=build --exclude=examples/**,testing/**
⌘K
fix(ai, ai-harness): a guard tier for tool calls, complete plugin removal, media and body limits, signing keys, and turn and config fixes (#1824)
Succeeded
in
48s
Started
Linux
4 CPU cores
read-write
access token used
Copy command
nx run-many --targets=build --exclude=examples/**,testing/**
1823
404d417
GitHub
Open on GitHub
Tasks
Resource Usage
ta
TanStack
OSS