vinta-ai-workflows 0.7.0-alpha12 → 0.7.0-alpha14

package/CHANGELOG.md

@@ -7,6 +7,8 @@ the project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [0.7.0] — YYYY-MM-DD
 
+<!-- pre-release: 0.7.0-alpha14 on 2026-09-16 -->
+<!-- pre-release: 0.7.0-alpha13 on 2026-09-16 -->
 <!-- pre-release: 0.7.0-alpha12 on 2026-09-15 -->
 <!-- pre-release: 0.7.0-alpha11 on 2026-09-15 -->
 <!-- pre-release: 0.7.0-alpha10 on 2026-09-15 -->
@@ -22,6 +24,18 @@ the project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ### Added
 
+- **`vinta-ai-maestro gate <gate-id>` runs a declared gate from inside an agent
+  turn.** A phase used to run its full gate suite five or six times a round —
+  implementer, reviewer, fixer, reviewer again, then the authoritative `gate`
+  node — and only the last of those consulted the gate cache or the resource
+  pool. Agents now ask the daemon for a gate by id: the result is cached on the
+  same `(gate id, lane tree hash)` key the `gate` node reads, the gate's
+  resources are held by the daemon rather than by whichever agent remembered to
+  wrap the command, and the id resolves to the plan's own command so a scoped
+  approximation cannot be reported as the gate. Each run is journalled as a
+  `gate_result` event and a `gate_run` transcript entry under the GATE band,
+  whoever asked for it.
+
 - **A project declares what makes a lane runnable.** `project.env_files` copies
   ignored configuration into each worktree, `project.setup_cmd` runs an
   idempotent setup hook on provisioning and recycle, and the fixed
@@ -184,10 +198,33 @@ the project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
     claimed per review turn rather than per phase — it has one session ledger and
     two concurrent reviews would collide over it — so one reviewer on a three-lane
     plan is a queue at the review step and not a serialised run.
+    - **"Cheapest qualified" has one exception, and it is about sessions.** A
+      member who already holds a session this phase would genuinely resume takes
+      it ahead of a cheaper member who would have to start cold — even when they
+      are more senior than the phase needs, and even when the member the plan
+      named is free. What it buys is the context a cold session spends its first
+      turn rebuilding; what it costs is the dearer model for that phase, which is
+      the smaller of the two bills. **The tier floor is untouched**: warmth
+      reorders the members who already qualify and never widens them, so a warm
+      junior still cannot take a Tier 3 phase.
+    - **Only when a session is genuinely already up.** "Warm" means the scheduler
+      asked its own session-reuse rules and got a yes — same pinned lane, same
+      harness, under the turn ceiling, not poisoned by a failed phase. A member
+      who has merely run before is not warm, because a promotion bought on a
+      resume that is then refused pays the senior's rate *and* cold-starts
+      anyway. With nobody warm, staffing is exactly what it was: a session is
+      being opened either way, and the plan's own level wins.
+    - An operator who picks the member by hand on a retry gets that member. A
+      saving nobody asked for does not overrule an explicit answer.
   - The run view reports who actually worked against who the plan said would.
     `substituted` is the figure to read beside a cost that overran: every
     substitution ran at or above the budgeted tier, so a run can be entirely green
-    and still have been staffed dearer than planned.
+    and still have been staffed dearer than planned. The phases promoted to reuse
+    a warm session are counted **within** that figure and reported apart from it —
+    a peer covering for a busy member costs what the plan budgeted, while a warm
+    promotion is the scheduler deliberately trading model rate against cold
+    starts, and folded together a run that promoted everything to the top tier
+    would read as an ordinary busy wave.
 
 - **`implement-plan` runs independent phases in parallel.** The plan now carries a
   dependency graph and the conductor schedules against it: a phase starts as soon as
@@ -282,8 +319,63 @@ the project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   **generated** from a zod source rather than hand-written — see
   [`schemas/README.md`](schemas/README.md).
 
+- **The monitor keeps thinking when you look away, and thinks out loud.**
+  Asking used to be one long HTTP request that produced the whole answer inside
+  it, which made the turn's lifetime the browser's connection: switching view,
+  reloading or letting a laptop sleep killed the monitor mid-thought and left a
+  question with no answer and no record that one had been attempted. The daemon
+  owns the turn now — `POST` answers `202`, and the monitor journals each
+  thought and each sentence as it produces them, so the conversation is
+  readable while it is being written and by any tab that opens it. The panel
+  renders through the same rows a phase's transcript uses, which is what the
+  API always claimed and the UI never did.
+
+- **A phase's transcript says which agent wrote each line, and holds the gates.**
+  Every spawn on a node appends to one file whatever its role, so a phase that
+  took two fix rounds held five agents' output in one undifferentiated stream —
+  the role was in scope at the append and simply not written down. Each line now
+  carries the role and the session slot that produced it, the view bands the
+  list where one agent stops and the next starts, and gate runs are entries in
+  it rather than only in a log beside it. The operator's own steering stays the
+  operator's (§7), even though the adapter echoes it back on the agent's event
+  stream. Transcripts written before this render exactly as they did, with no
+  bands. The band stays pinned to the top of the scroller while its own run of
+  rows goes past, so the agent you are reading is named without scrolling back
+  to where it started — in the phase transcript and in the monitor conversation
+  alike.
+
 ### Changed
 
+- **A session cannot inherit a machine's decision to stop compacting.** All
+  three harnesses compact automatically when their context fills — there was
+  never a flag to turn on — but each honours an environment variable that turns
+  it off, and an agent spawned on a machine carrying one would run a long phase
+  straight into a wall it was supposed to be able to survive. Those variables
+  are now stripped from every child the daemon spawns, and claude-code's own
+  setting is reasserted in the per-lane file beside its permission rules. The
+  window itself is left alone: an operator who narrowed theirs compacts earlier,
+  which is not the failure this guards against. A harness that compacts declares
+  it, and a compaction the harness reports is a transcript row like any other.
+
+  One consequence worth knowing: `max_session_turns` used to be a proxy for "this
+  session is nearing a context window it will die on". That death no longer
+  happens, so the ceiling now guards a different thing — a session whose memory
+  was replaced by a summary of itself, and which answers just as confidently.
+
+- **Agent prompts name gates by id rather than printing their commands.** The
+  implementer's step 4 was descriptive — "Outer gate… These run against your
+  lane:" followed by the gate commands — and implementers were observed reading
+  it as a note about what would run later, finishing the scoped inner loop and
+  never running a gate at all. It is imperative now, and every role that runs a
+  gate (implementer, reviewer, fixer, and the continuation of each) is pointed
+  at `vinta-ai-maestro gate <id>`. The fixer prompts no longer say to re-run the
+  gates "until they are green", which contradicted the commit protocol directly
+  below them and left fixers running out the turn with the work uncommitted;
+  they now say to commit and report FAILURE when a gate cannot be turned green.
+  Implementers and fixers report which gates they ran and what each returned —
+  a record for the reviewer to check against, explicitly not a licence for the
+  reviewer to skip running them itself.
+
 - **Phase branches base on their dependencies, not on the previous phase.** A phase
   with no dependencies cuts from the default branch; one with a single dependency
   cuts from that phase's branch; one with several cuts from a
@@ -337,8 +429,110 @@ the project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   is only as good as the `healthcheck` a service declares: without one, compose
   returns as soon as the container starts and `createdb` races the server.
 
+- **A merge conflict goes to the most senior member who wrote one of its
+  sides.** The conflict fixer was staffed from `defaults` and ignored the roster
+  entirely, so on a tiered crew a collision between two Tier 4 phases was handed
+  to a default-tier agent that had read neither of them — the hardest merge in
+  the run going to the least contextualised agent available, and resolved by
+  guessing which side looked more finished.
+
+  `Integrator` already knows which nodes own the contested paths, and `node_crew`
+  already records who took each one; this is the join. Highest tier among them
+  wins, on that member's own `model` and `harness`, and a tier tie breaks on
+  member id the way roster substitution breaks its own. An unstaffed workflow
+  still resolves through `defaults`, unchanged, and so does a conflict whose
+  nodes cannot be resolved to a member.
+
+  What is carried is the member's **capability, not their session**. A phase's
+  session ran in that phase's lane; the fixer runs in the integration worktree,
+  where the files in dispute are half-merged and unlike anything that session
+  saw — resuming there would give the agent confident, false memories of exactly
+  the files it is merging. The prompt tells the fixer which side is its own, and
+  in the same breath that it is a fresh session and must read rather than
+  recall.
+
 ### Fixed
 
+- **An agent taking a lease left the resource panel showing the one before it.**
+  The scheduler's own gate-pool transitions are journalled as `gate_pool`
+  events, so the browser re-reads the snapshot when one lands. An agent's lease,
+  taken through `vinta-ai-maestro with`, was only a row in the `leases` table —
+  correct in the daemon's projection and announced to nobody. A projection
+  nobody is told changed is a projection nobody sees change, so the panel showed
+  whatever holders it had last time some unrelated event happened to arrive, and
+  the waits it was stalest about were the agent ones, which are the long ones.
+  The broker now journals `agent_lease` on both edges. A renewal writes nothing:
+  it is a heartbeat on a lease already reported, and journalling it would bury
+  the two rows that are transitions under a liveness log.
+
+- **An agent waiting for a semaphore could be overtaken without bound.** A
+  `vinta-ai-maestro with` wait is served in fifteen-second hops so no client
+  timeout has an opinion about it, and each hop used to abort the underlying
+  `pools.acquire`. Aborting leaves the pool queue, and the next POST re-entered
+  it — at the tail, behind every waiter that had arrived in the meantime. Since
+  a gate run holds a capacity-1 semaphore for minutes, an agent lost that race
+  on hop after hop while the scheduler's own in-process waiters, which never
+  leave the queue, did not. The pool's aging rule could not help: aging reserves
+  against waiters *behind* the aged one, and the waiters that overtook it are in
+  front of it. A `202` now carries a `waitToken` naming the wait, the client
+  sends it back, and the daemon keeps the one queue entry — one arrival time,
+  one place in line — across the whole wait. A parked wait whose client stops
+  hopping is reaped after a minute, releasing the slot if one was granted to it
+  while nobody was watching.
+
+- **The monitor's conversation could never be written on Windows.** It is kept
+  under a reserved node id that becomes a directory, and that id was
+  `monitor:conversation` — a colon, chosen because no phase id may contain one,
+  and the drive separator on Windows. `mkdir` failed with `ENOENT`, so every
+  question threw on its first append and the endpoint reported the monitor
+  unavailable. A colon is legal in a macOS or Linux filename, which is why no
+  machine the feature was built on ever saw it. The id is now
+  `_monitor-conversation`: still unrepresentable as a phase id, and legal
+  everywhere. Conversations recorded under the old id are not migrated.
+
+- **An implementer was told both to commit and never to commit.** The prompt
+  carried "never commit while a gate is red" beside "the turn is not complete
+  until `git status --porcelain` is empty of your work" — a contradiction
+  whenever the gate could not be turned green inside the turn, which is most of
+  why fix rounds exist. Agents obeyed the `never`, and the reviewer one node
+  later raised the BLOCKER it is told to raise for uncommitted work, spending a
+  fix round re-implementing code that was already on disk. A red gate now
+  changes the report and never the decision to commit: commit, and report
+  FAILURE saying what is still red.
+
+  The reviewer's side was a false BLOCKER independently of that. Any unclean
+  tree was a finding, and a lane's tree is never clean — the pool copies
+  configuration in, links dependency trees, and the gates leave build output
+  behind, none of which the implementer is allowed to stage. It is scoped to
+  the failure it was written for now: work that is in the tree and missing from
+  the diff.
+- **A live transcript stops following at entry sixty.** The effect that kept the
+  box on the newest row was keyed on how many rows were *shown*, which is
+  `min(entries, window)` — so it stopped changing the moment a transcript
+  outgrew one window, and the following died silently and permanently, including
+  after the operator scrolled back to the bottom to ask for it. Which is roughly
+  minute two of any real phase. It follows on the entry count now, growing the
+  window keeps the reader on the row they were reading instead of teleporting
+  them, and a **Jump to latest** button says so out loud rather than leaving a
+  60px band at the bottom of a scroller as the only way back in.
+
+- **The transcript spends its space on the noisiest rows.** An agent's answer,
+  its private reasoning and four hundred characters of tool-call JSON all
+  rendered as the same two lines at the same weight. Thinking is now small and
+  quiet, and a streamed thought is one row rather than the dozen events it
+  arrived as; tool calls and their results fold to the argument that says what
+  they did — a command, a path — and open individually or a whole kind at a time
+  from the panel header. Prose is untouched and never folds.
+
+- **A panel's expand control did nothing on the node view.** It expanded by
+  taking `col-span-full`, which on the run view bought a strip a third wider and
+  on the node view matched nothing at all — those panels sit in a flex column,
+  not a grid, so the only thing the button changed was the height of the
+  scroller inside it. It now opens the panel over the whole page through a
+  portal, so no scrolling ancestor can clip it, with Escape to leave and the
+  page behind it held still. A transcript that was following stays on its newest
+  row across the change.
+
 - **`doctor` checks that the shared servers answer.** It verified that binaries
   existed and disk fitted, and never looked at a `server_url` — so a project
   whose Postgres lives in another checkout's compose stack passed the preflight
@@ -1178,7 +1372,6 @@ the project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
   **Consumers**: re-sync to pick up the opt-in questions. E2E-disabled projects
   are unaffected (the e2e regions strip out as before).
 
-
 <!-- pre-release: 0.2.0-alpha6 on 2026-07-13 -->
 
 ### Changed
@@ -1215,7 +1408,6 @@ the project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ### Added
 
-
 - **`handoff` foundation skill — session-continuation handoff docs between
   agents.** Write mode captures the current task (goal, verified-vs-unverified
   state, decisions + rejected alternatives, landmines, single concrete next

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vinta-ai-workflows",
-  "version": "0.7.0-alpha12",
+  "version": "0.7.0-alpha14",
   "description": "Bootstrap AI tooling (AGENTS.md, skills, sub-agents) across Claude Code, Codex, Cursor, and VS Code + GitHub Copilot.",
   "type": "module",
   "bin": {