Repository navigation
Conversation
…inding fills (#3186) A metric/status/progress widget bound to a declared metric gets its value (and a status widget its color) from the recorded points on every read (ent#479), and the agent server's validate_widget already exempts them. D-003 still demanded value/color, so an agent following the documented binding contract was reported HARD-incompatible. D-003 now applies the agent server's predicate (a truthy `metric`) and skips exactly the filled fields (_BOUND_FILLED); `label` stays required and unbound widgets keep today's rule. Decided: a binding to an undeclared name is exempt too — the backend renders "not declared" and drops any written value, so requiring one demands a number nobody sees. D-009 checks the metrics block's shape only; no check flags undeclared bindings yet. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…#3214) Docker volumes are daemon-global while the sweep's ownership check reads one stack's DB, so a second stack on a shared daemon reclaimed every other stack's unattached agent volumes — force-removed, logged at INFO. - Every agent data volume is created through docker_utils. agent_volume_labels, which stamps trinity.instance=<installation_id> (instance_identity.get_instance_id — the durable id the alert label already abbreviates; one source). All six creation sites use it, pinned by a guard test. - The sweep lists candidates on the instance label's key AND value and skips the cycle when the id is unresolvable. - is_reclaimable_agent_volume refuses a volume labelled for another instance, closing the same hole on the retention-purge path (two stacks can both have an agent `alpha`). - Legacy (unlabelled) volumes fail closed: labels are immutable, so the sweep never reclaims them; unowned+unattached ones are named in one WARNING per change of that set for a human. Own agents' volumes are still removed at retention purge, which is ownership-row driven. - Reclaims and removals log at WARNING as unrecoverable, naming volumes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… — never re-send on silence (Abilityai/trinity-enterprise#568) The receipts already existed on every dispatch route (#914, #2661, #2670); what was missing was the caller knowing what they mean. In the 2026-09-08 cascade every duplicate dispatch was an agent re-sending after "could not confirm delivery". The contract is now ONE text, taught wherever a caller reads: - the platform prompt, §Agent Collaboration (DELEGATION_CONTRACT spliced into PLATFORM_INSTRUCTIONS) — every agent, every turn, no image rebuild; - verbatim in the chat_with_agent and every dynamic chat_with_<agent> description (src/mcp-server/src/delegation_contract.ts) — the copy an external MCP client, or an agent at PromptTier.MINIMAL, reads; fan_out and send_message repeat its rule sentence and point at it; - the receipt message itself: every MCP-authored receipt ends with "Do not re-send: read the outcome with get_execution_result(agent_name=…, execution_id=…)", chat() reuses queuedTimeoutReceipt, and runAgentChat rewrites the backend's REST-only "Poll GET …" line on accepted/queued. Claude Code cuts every MCP tool description at 2,048 characters (CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH). The old chat_with_agent description was 2,424, so its async and list_recent_executions advice never reached a model. It is now lead + modes + contract (~1.9 KB); per-mode detail moved into parameter descriptions, which are not cut. The text names only what the platform produces today: no `replayed` (ent#566 has not shipped it), the agent.task.* wake scoped to parallel runs (a sequential /chat turn emits none), `retryable: false` → "do what its message says" (gate refusals say "try again shortly"). The unverified "concurrent-duplicate guard will kill mid-execution" claim is gone from the description, the receipts and requirements §37.1 — no such guard exists. Tests: tests/unit/test_ent568_delegation_contract.py (byte parity with the TS copy via a strict JSON parse + guard-the-guard cases; every tool, argument, status and event it names proven real; budget; AC4) and src/mcp-server/src/delegation-contract.test.ts (the descriptions a real tools/list publishes, under the cap; receipt messages on chat/task timeout, 409 replay, async accepted/queued and the pull-routed path). 15 wiring mutations each go red. CI boot smoke also imports dist/tools/dynamic-agents.js. Docs: requirements §37.5, architecture/mcp-server.md, six feature flows, the agent guides' stale "will fail after 60 seconds" sections, user-docs. Follow-ups filed: #3232, #3233, #3234. Implements Abilityai/trinity-enterprise#568. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…mport - test_2669: get_instance_id() mints installation_id on purpose. A volume created before the id exists would carry no owner and never be reclaimed. It gets an _ALLOWED_USES entry with that reason, like the label tier's existing entry. - test_agent_readiness_probe: the hand-written services.docker_utils stub lacked agent_volume_labels, which lifecycle.py now imports. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e's 2,048-char description cap (#3234) Claude Code shows a model only the first 2,048 characters of an MCP tool description. The server publishes the full text, so the cut tails were invisible from Trinity's side. - deploy_local_agent (2,698 -> 1,477): the four packaging steps, including the manifest script, move into the `archive` parameter, whose description is not cut. Steps 1-2 are byte-identical to before. The description keeps the rules up front. - get_objectives (2,349 -> 2,034): this tool has no parameter to carry detail, so the wording is tightened instead with no instruction dropped. "stale: DO NOT ACT" now comes right after the purpose. - tool-description-budget.test.ts: a real createServer, a user-scoped key and tools/list over streamable HTTP. It fails naming every published tool over the cap with its length. chat_with_agent (2,424) is a shrink-only PENDING entry owned by Abilityai/trinity-enterprise#568, which is rewriting it; the entry fails once that tool fits. Fixes #3234 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… out (Abilityai/trinity-enterprise#568) AC1 lists `replayed` among the receipts. No dispatch route emits it until ent#566 ships intent-scoped dedupe (only ask_operator answers `replayed`), and the guard that checks every named status against a dispatch-route producer would refuse it. A replay is still covered by "whatever its status" and "an exact repeat is normally answered with the original" — observed live on localhost: an identical repeat answered in 28 ms with the first call's execution_id, and one task ran. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rding, history corrected (Abilityai/trinity-enterprise#568) The PR #3240 review found three places where the change said more than the code does. - Contract, bullet 4: `agent_busy` is not proof that nothing ran. The SUB-003 usage-limit 429 is raised after the turn ran, and the MCP client labels it `agent_busy` (#3244). The bullet now sends that case through the same `list_recent_executions` check as every other error without an `execution_id`, then a word-for-word re-send after `retry_after_seconds`. - Receipt leads: the backend stores the async receipt as the idempotency snapshot at dispatch time and replays it for 24 h (#3245), so "Accepted: the task is running" can be false on a replay. The leads now claim nothing about liveness. A replay of the backend's own sync-/task `queued_timeout` snapshot, which tells a REST caller to poll GET, is rewritten too; the MCP server's own `queued_timeout`, which already carries the read line, is left as it is. - History: Claude Code's 2,048-character description cap has cut the `chat_with_agent` description since #2958 (2026-09-24), which took it from 2,035 to 2,424 characters. The receipt advice was delivered before that. Requirements, the learnings fragment and the CSO report now say so. Tests: every `chat_with_*` tool must take `parallel` and `async`; the dynamic-tool factory keeps the backend's line and falls back to the exact default; each lead is pinned per status for agent, user and self-task sessions; the backend `queued_timeout` replay is rewritten and the MCP server's own is not. Ten mutations each turn a test red: a dropped `async`, swapped or optimistic leads, a skipped rewrite per session kind, a factory that ignores the backend line, both `queued_timeout` branches, and TS-only contract drift. Docs: execution.md names the sync `/chat` exception (it emits no `agent.task.*` event), requirements §37.5 says why `replayed` is left out, and follow-ups #3244 and #3245 are linked. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Abilityai/trinity-enterprise#568) The history now says Claude Code cut the `chat_with_agent` tail before the model read it, instead of "any model" or "every agent": the 2,048 cap was read out of the Claude Code binary, and agents on the Codex or Gemini runtimes reach the same tools through MCP clients that were not measured. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This was referenced Oct 5, 2026
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Integration surface for #3238, #3230, #3231, #3240. Never merged; members merge individually once green.
Carries one train-only commit: the
chat_with_agententry is removed fromPENDINGintool-description-budget.test.ts, because #3240 brings that description under the cap and #3238's test fails an entry that fits. The same removal goes onto #3240's branch after #3238 lands.