Skip to content

v4.0.1

Choose a tag to compare

@github-actions github-actions released this 11 Nov 14:58
· 11 commits to main since this release
bc0f7e1

4.0.1 is a critical fix. The run token and cost budgets shipped incomplete in 4.0.0: usage could not be reported once a run's members were terminal, and the budget measured the last turn's figure instead of the run's spend. They are removed (see Removed); a client that still sends a removed field gets unknown_parameter. There is no migration, so a rollback to 4.0.0 is safe.

Added

  • runs_read frontier reports cancelled members and the in-flight items' titles. cancelled counts the terminal members that carry the status label of cancel under the run's root (the one label policy: per-root status_labels, then the server's, then the built-in cancelled); they stay counted in counts.terminal. Each inFlight entry gains its item's title. Both are additive, on MCP and on POST /api/v1/ops/runs_read.

Plugin

  • Fixed: run-wave post-run reports name the pause decision and count open decisions. report --event paused takes --paused <file>, a saved run_paused answer (a write answer, a top-level error or a bare {code, detail}), and prints its pauseDecisionId and pauseReason; the run view never carried them. A --paused file with no run_paused error exits 2. The closed report reads the close snapshot's openDecisions as the integer the server sends. A reviewer assign that hits run_at_capacity backs off instead of routing to Step 5a, and the plugin prompts, ralph-loop and the run-wave, director and schema-workflow docs drop the removed paused: {decisionId} write-success member and the amount answer field. The work-summary runs section and the api-reference report row no longer describe run budgets.
  • Changed: the director guard covers host MCP tools, the web tools and the remaining body reads. Under the director marker, every mcp__ tool whose name lacks task-orchestrator (terminal, file-store and computer-use servers, or a Task Orchestrator server registered under a key without that word) is denied unless its bare name is start_session, send_message, update_plan, wrap_up, mark_chapter, get_settings or set_session_output_style, through a third guard registration (hostToolMatcher()). WebFetch, WebSearch, ReadMcpResourceTool and ReadMcpResourceDirTool are denied; config is denied and runs_read is allowed only for get, list, frontier, decisions and decision; Skill runs only the director skill and task-orchestrator:model-effort; gh pr view and gh pr list take only operands, -R, listing filters and --json with status, metadata and body fields (no comments or reviews); a Workflow call carrying scriptPath or resumeFromRunId is denied, so a tier-2 resume goes to the coordinator. Call and docs updated in the director and run-wave skills and the hooks README.
  • Removed: every usage-report path is removed with the server's heartbeat usage and run token and cost budgets. run-wave post-run no longer writes usage.json or sends a usage heartbeat (Step 0a and the Step 7 heartbeat are gone), run-ops --phase heartbeat and --usage are removed, and run-ops --phase open sends only attemptsPerSeat and maxActiveSeats as the run budget. plan warns once per key when a profile sets runBudget.tokens or runBudget.costUsd (ignored; maxActiveSeats still applies). ralph-loop drops --run-budget-tokens and --run-budget-usd (now exit 64, like --actor) and no longer reports usage before the complete transition; the ralph, run-wave, director, init and manage-schemas docs no longer describe token or cost budgets (#677).
  • Fixed: the orchestration-mode band shows only this session's runs, in every mode. It drew every open and paused run on the server, and counted every session's decisions, so a session with nothing of its own showed another session's run, even one from another project. It now draws only the run the session coordinates and the runs it or one of its subagents (a background coordinator, a Workflow, a seat) opened or acted on through runs_write, and no longer reads the server-wide decisions list. Touching or reading an item claims no run, so another session's run over the same feature stays out. Only a write moves the focus: an items_read naming an item (a status check, a look at another session's work) no longer makes it the session's focus, and a focus a read set is dropped on the next load. Director mode follows the same rule: the band is the session's own graph, not the board.
  • Fixed: the orchestration-mode band no longer reports a stalled coordinator for a Workflow-driven run. The alert follows the director skill's rule: nothing while a Workflow runs the run (hints.method A, hints.phase launched), while any seat is live, or on a paused run. Before, any Workflow wave longer than the stall threshold showed COORD STALLED until the post-run step.
  • Changed: the orchestration-mode band speaks plain words and shows each agent. The title row carries the bar on its right: one cell per member, green done, a red ⧄ cancelled, red stuck, amber waiting on a decision, grey finished and waiting to be advanced, hollow not started, and each member an agent works on in a colour of its own that it keeps from stage to stage, blinking. The status row puts what needs a look first (1 agent stuck, Coordinator not responding, 1 decision for you, Run paused, All agents finished · advance the items), then what the agents do (3 agents implementing), with the members by phase in plain words on the right (3 planned · 2 in progress · 1 done); the run ids, ready/waiting and q → w → r → ✓ are gone. Below it, one row per agent (its marker in the member's colour, blinking with its bar cell, then its seat and the item's title; at most three, then +N more), else the item the session last touched. The blink runs in a Client on the surface's own clock, one timer for the whole band, with no extra server reads; VS Code and the mobile app draw the band still. A short band drops rows from the bottom, never the title. The bar and the counts never shrink: a long title gives way (on the desktop app it wraps onto a second line) instead of squeezing the bar.
  • Fixed: the orchestration-mode band's focus stops below a backlog bucket. With no run, the band showed the ancestor just below the project root of the item the session touched, so a task in a bucket such as Bugs drew the bucket's whole backlog, and counted the bucket itself as in progress. The walk now stops below a grouping item (typed container or project, or tagged container), so the band shows the task's own feature; a bucket the session touched itself counts its work, not itself. The stored focus is now a cache of the touched item's walk, walked again once per load, so a focus stored before a reload or a resume follows the current rule.
  • Fixed: run-wave provenance records each seat's transcript meters instead of the run notification figure. run-planner provenance now reads every seat's input, cache write, cache read and output meters from dispatch-meta for both methods and writes one measure on every line: tokens is the non-cached sum, with meters= and seat-meters= tails and six-field extra-seats rows carrying the four meters. see:<short> and the Workflow notification figure are gone, duration reads unknown, a seat without a transcript reads unknown with a stderr warning (never zero), --usage is ignored with a warning, and dispatch-meta also matches planner-a<n> and reviewer-a<n> labels. Notes written before this change still parse.
  • Fixed: run-wave seats that own implementation-notes or session-tracking now receive the closure-notes rule. The run planner derived a stage's rules only from the skill: of the notes it owns, and no note names closure-notes, so a generic workflow seat (no dispatch.agent) never got the rule. deriveRulesSkills now adds it once whenever the server serves it (#676).
  • Fixed: the run-wave fix-cycle reseat prompt tells the seat to rewrite its notes in full. A fix-cycle seat (-a<n>) kept the first seat's banned sections and appended a per-cycle narrative to its notes. The FIX CYCLE line now ends with a rewrite-in-full instruction for each of implementation-notes and test-manifest the base stage owns: replace superseded text, drop the sections the closure-notes rule bans, add no per-cycle section, and keep history to a final Amendments: list.
  • Bumped plugin versions: task-orchestrator to 4.0.1 (the usage-report removal and the run-wave and director fixes above) and orchestration-mode to 0.1.1 (the band fixes and redesign above).

Removed

  • Run token and cost budgets and agent-reported usage are removed from the MCP and REST surface and from config (#677, #678). They shipped incomplete in 4.0.0: usage could not land once the members were terminal, and the budget measured the last turn's figure instead of the run's spend. Removed: advance heartbeat usage; runs_write open / update budget.tokens and budget.costUsd; runs_write answer amount; the run view's usage.reportedTokens / usage.reportedCostUsd and the budget.tokens / budget.costUsd members; the usage member of a runs_read seats row; the paused: {decisionId} member of write results; the budget section of the report run rollup; and tokens and costMicroUsd in the budget payload of the run.opened event (it now carries the attempt and seat caps only). A request that still sends a removed field fails unknown_parameter on that field (/usage, /budget/tokens, /budget/costUsd, /amount), so a hardcoded REST client must drop it. runs_read get keeps usage.observedTokens, the server's own estimate, which enforces nothing. The attempt cap (attemptsPerSeat), the seat cap (maxActiveSeats) and the attempt-cap pause and decision are unchanged. A config document that still sets runs.budgets.tokens or runs.budgets.cost_usd loads: it validates and pushes (any value), each key draws one warning at its pointer, and no ceiling is applied. There is no migration: the runs and run_seats columns stay and are no longer written, so a rollback to 4.0.0 is safe. State 4.0.0 stored still works: a run paused by a run-budget decision resumes on extend (no amount, no budget change) and closes budget_exhausted on stop, and a stored rollup snapshot with a budget section decodes with it ignored.

Fixed

  • runs_write assign accepts seat reviewer on a seat-less item whose schema has a review-role note. A seat-less schema (no seats: after trait merge) allowed only implementer and planner, so assigning a reviewer was not_found and review-wave could not run a token-bound review for such items (this repo's feature-implementation). The allowed set is now those two plus reviewer when the resolved schema has any review-role note; a seat-less schema without one, a schema-free item and a seat-aware schema that does not declare reviewer still answer not_found. No lane is enforced on a seat-less schema: the first writer's seat owns a note.