v0.7.0

Added

• Demo GIF + scripted replay. New demo/replay.mjs (a presentation-only
  replay of a /maestro task run that closely mirrors the runtime's real log
  output) and demo/demo.tape (VHS scene) render docs/assets/demo.gif,
  now embedded at the top of both READMEs. Dev-only: demo/ and
  docs/assets/ are excluded from release archives.
• Always-on token accounting. Completed task runs log their total token
  usage (tokens=<used>[/limit]) and persist tokensUsed in the run
  manifest (budget-stopped runs too); /maestros shows tokens=<n> per run.
  Enforcement stays opt-in via GHCP_MAESTRO_BUDGET_TOKENS — accounting
  always-on, spend caps only when explicitly configured.
• Cross-agent prompt-injection hardening. Dependency outputs
  (augmentPromptWithDeps) and the synth digest (buildSynthPrompt) are now
  fenced in <<<UNTRUSTED-AGENT-OUTPUT>>> sentinels with an explicit
  treat-as-data instruction, per OWASP Agentic Security guidance — agent
  output is data, not instructions. Always-on, zero token cost. (#33)
• Opt-in per-label model routing. New core/model-routes.mjs:
  GHCP_MAESTRO_MODEL_ROUTES (or opts.modelRoutes) maps label patterns
  (plan, explore:<agent>, verify, synth; * wildcards,
  first-match-wins) to models so fan-out workers can run on cheaper models
  than the planner/synth. No routes → adapter default, specs unchanged. (#17)
• Opt-in verify phase. GHCP_MAESTRO_VERIFY=1 (or opts.verify) inserts
  a verification agent between explore and synth that judges each subtask
  output against the original task objective (met / partially-met /
  not-met + gaps, MAST-style); the report feeds the synth prompt. Verify
  failure is non-fatal. Off by default — one extra agent per run. (#31)
• OTel GenAI-style trace export. New core/trace.mjs + writeRunTrace
  in run-flow.mjs: every terminal run (complete / stopped / error) writes
  trace.json — an invoke_workflow root span plus invoke_agent child
  spans with gen_ai.* semantic-convention attributes. Always-on, zero deps,
  best-effort (never fails a run). (#32)
• Release polish. New docs/DEMO.md — a five-minute end-to-end
  /maestro task walkthrough with real log-line formats (install → plan gate
  → parallel explore + /maestros dashboard → synth → resume/stop). READMEs
  gain a Known limitations section, the copilot plugin install flow was
  verified end-to-end against the self-served marketplace (0.5.0 → 0.6.0
  upgrade), and the release-archive file list was re-confirmed lean after the
  docs reorg. (#24)
• /maestro install <source> [--force] — one-command workflow install.
  New core/workflow-install.mjs: accepts a github.com blob URL, a
  raw.githubusercontent.com URL, or owner/repo/path/flow.mjs[@ref] shorthand
  (ref defaults to main; https only), downloads the file (256 KB cap, checked
  against Content-Length before the body is read; redirects refused so a raw
  URL cannot bounce off GitHub), and validates it without executing it —
  node --check parse + default/run export scan on a quarantined temp copy —
  before writing to the user workflow dir; project workflows still shadow
  it. Existing names are never overwritten without --force, installs confirm
  via elicitation when the host supports it (dialog errors fail closed), and
  every install prints a review-before-running warning. install joins the
  reserved workflow names. (#23)
• Agent auto-retry with exponential backoff. spawn/spawnAll retry
  error-status attempts (API blips, rate limits) with exponential backoff +
  jitter — GHCP_MAESTRO_RETRIES extra attempts (default 1, 0 disables).
  timeout/aborted outcomes never retry, backoff sleeps abort with the run
  signal, and results/persisted agent records carry an attempts count. The
  quality helpers forward retries/retryBaseMs. (#20)
• Cost visibility + token budget cap. New core/budget.mjs: the task
  workflow logs a run-size estimate at the plan gate (also shown in the
  approval dialog), warns on fan-outs of GHCP_MAESTRO_LARGE_RUN_AGENTS+
  subtasks (default 5), and GHCP_MAESTRO_BUDGET_TOKENS (accepts 500k/2m)
  soft-stops a run once exceeded — in-flight agents finish, un-started agents
  are skipped as aborted, the run is marked stopped and stays resumable
  via /maestro-resume. spawn/spawnAll/runPhase accept a shared
  budget tracker. (#14)
• DAG plans: dependsOn between subtasks. The plan schema gains an
  optional dependsOn: string[] (validated: known agents, no self-reference,
  no cycles). planLayers groups specs into topological layers; the task
  workflow fans each layer out in order and injects dependency outputs
  (truncated to 4 000 chars each) into dependents' prompts via
  augmentPromptWithDeps. Dependents of a failed dependency are persisted as
  a new skipped status without invoking the adapter, so /maestro-resume
  reruns them. Plans without dependsOn behave exactly as before. (#21)
• Partial-failure disclosure in synthesis. agentDigest renders non-ok
  results as ## <agent> (FAILED: <status>) blocks with the error text;
  buildSynthPrompt adds a "state explicitly which angles are missing"
  instruction when any subagent failed (all-ok prompts are byte-identical to
  before); the task workflow logs a coverage: N/M subtasks ok (…) line next
  to the final answer. Both surfaces share the change via core/synth.mjs. (#22)

Fixed

• Resume now reruns failed agents. spawn replays only cached ok records
  from the run store; cached error/timeout/aborted records re-invoke the
  adapter on /maestro-resume, matching the documented resume contract
  (previously a failed agent's cached failure was replayed forever).
• In-flight cancellation actually interrupts agents. The standalone-client
  adapter races sendAndWait against the caller's AbortSignal (raceAbort),
  so timeouts and user cancellation no longer wait for the model to finish.

Changed

• READMEs slimmed, docs split. The long "Features" section and the
  configuration table moved to docs/GUIDE.md / docs/GUIDE.ko.md; the
  CLI-vs-VS-Code install-surface details moved to docs/SURFACES.md. The
  READMEs keep a compact feature summary, the commands table, and links.
• Design docs consolidated under docs/specs/. The docs/superpowers/
  tree (specs + implementation plans) is gone; all design/plan documents now
  live in docs/specs/ and the stale export-ignore rule was dropped.
• /maestro-stop aborts in-flight agents. New core/run-registry.mjs
  (process-local runId → AbortController); runPhase wires the registry
  signal by default and stopRun aborts it, so stopping a run started in the
  same session interrupts its agents instead of only flipping the manifest.
• /maestro workflows rescans the workflow dirs (and /maestro run
  rescans on a name miss), so workflows added after startup are picked up
  without restarting the CLI.
• VS Code runtime bridge deduplicated against core: uses
  runWithConcurrency and DEFAULT_CONCURRENCY from core instead of local
  copies; the <name> [args] split is shared as
  splitWorkflowInvocation in core/saved-workflows.mjs.

v0.6.0

Adds a VS Code surface alongside the CLI plugin and extracts the runtime into a
shared, zero-deps core/ package consumed by both. The CLI behaviour is
unchanged; the plugin now imports the same core the VS Code extension uses.

Added

• VS Code /maestro chat participant with a TUI-like Run Console (webview)
  and an Activity Bar Runs tree for per-agent drill-down (model, tokens, tools,
  timing, prompt/output/error, retry). Ships as a separate VS Code extension
  (vscode-extension/), not part of the CLI plugin archive.
• Shared core/ package (@ghcp-maestro/core): the fan-out, spawn,
  run-store, saved-workflows, quality, plan/approval, probes, and synth logic —
  surface-neutral and unit-tested — now imported by both the CLI plugin
  (extensions/ghcp-maestro/) and the VS Code extension.

Changed

• Unified synth + run commands. core/synth.mjs (buildSynthPrompt) is the
  single source for the synthesis prompt across both surfaces; showRuns /
  resumeRun / stopRun moved into core/run-commands.mjs. The CLI reproduces
  its prior synth prompt and run-listing output byte-for-byte.
• Slimmed the CLI extension God-file (extensions/ghcp-maestro/extension.mjs)
  by extracting run-phase, builtin-workflows, and run-command helpers into core.
• Leaner release archive: vscode-extension/, scripts/, and design docs
  under docs/superpowers/ are export-ignored from the CLI plugin archive.

Fixed

• Run Console webview hardening: escape U+2028/U+2029 in embedded state
  (JS line terminators inside <script>) and replace the script-src 'unsafe-inline' CSP with a per-render crypto nonce.
• Cancellation is honoured end-to-end: a stopped run skips the synth phase
  and finishes as stopped instead of being overridden with complete/error;
  aborted/stopped/cancelled agents count as terminal and get a distinct
  tree icon.
• Portability: scripts/rebuild-install-stable.ps1 derives the repo root
  from $PSScriptRoot and resolves the code CLI from PATH — no hard-coded,
  username-leaking absolute paths.

v0.5.0

First tagged release. Highlights since the initial scaffold: the M4 dynamic task
workflow (plan → fan-out → synth), M5 saved workflows, M6 quality helpers, the
M4.x plan pre-approval gate, background runs with on-demand /maestros
monitoring, env-configurable agent timeouts, and CI. The runtime is zero-deps.

Fixed

• /maestros <runId> now shows the plan and synth phases too. Previously
  only the explore fan-out reported progress, so during the (often minute-plus)
  plan decomposition — and the final synth — /maestros <runId> showed "no
  progress recorded yet" and the run looked stuck. A new runtime/phase-monitor.mjs
  (startPhaseMonitor) wraps every phase, single-agent or fan-out: it records a
  pending snapshot to progress.json the instant a phase begins and threads
  the planner/synth streaming (bytes/tokens/state) through, so the phase is
  visible from the moment it starts. Applied across task (plan → explore →
  synth), brainstorm, and hello. New unit tests:
  tests/phase-monitor.test.mjs.
• Task subagents no longer time out on research-heavy runs. The task
  workflow's plan/explore/synth phases (and the brainstorm lenses) run full
  research subagents — web fetches, repo analysis, many tool calls + extended
  reasoning — whose wall-clock routinely exceeds the old hard-coded 120 s limit,
  so every subtask could fail with status=timeout and produce no output.
  Timeouts are now two tiers: research agents default to 10 min
  (GHCP_MAESTRO_TIMEOUT_MS) — matching the ~600 s floor of a single LLM API
  call in the OpenAI/Anthropic SDKs — and the fixed-prompt diagnostics
  (pong/hello/env probes) stay short at 1 min
  (GHCP_MAESTRO_TIMEOUT_PROBE_MS) so a slow one still fails fast. When a task
  aborts on timeouts, the failure message points to GHCP_MAESTRO_TIMEOUT_MS.
  New unit tests: tests/timeouts.test.mjs.

Changed

• Diagnostics hidden from /maestro help. hello and pong are
  infrastructure smoke tests, not user features — they now carry a hidden flag
  and render under a separate "Diagnostics" section instead of the main
  subcommand list (the commands still work, and the GHCP_MAESTRO_PROBE_* env
  triggers and resume registry are unchanged). The brainstorm summary drops the
  "demo" framing — it is a real multi-lens feature. Help rendering moved to a
  pure, unit-tested runtime/help.mjs (renderMaestroHelp).

Added

• Background runs + /maestros monitoring (issue #2). /maestro task|brainstorm|run (and the hello diagnostic) now dispatch in the
  background — the handler returns
  immediately with a running in background — watch with /maestros <runId>
  pointer and the session stays free while the fan-out runs. Per-agent progress
  (state, elapsed, streamed bytes, per-phase token totals) is aggregated by
  runtime/monitor.mjs and persisted to the run dir as progress.json. Watch it
  on demand: /maestros lists runs with a one-line progress summary for any still
  running, and /maestros <runId> prints the full dashboard. The standalone
  adapter subscribes to each child session's events (subagent.*,
  tool.execution_*, assistant.streaming_delta, assistant.usage) through an
  onProgress sink threaded by spawn/spawnAll. Monitoring is best-effort and
  never affects fan-out results; opt out with GHCP_MAESTRO_NO_MONITOR=1.
  New/updated unit tests across tests/monitor.test.mjs,
  tests/run-store.test.mjs, tests/spawn-progress.test.mjs,
  tests/standalone-progress.test.mjs, tests/monitor-enabled.test.mjs.
• M4.x — plan pre-approval gate. New runtime/plan-approval.mjs adds a
  planApprovalGate that runs after the plan agent decomposes a task but
  before the fan-out. On an interactive host (session.capabilities.ui. elicitation === true) it logs the subtask list + prompt previews and shows a
  multi-select elicitation dialog so the user can approve, run only a subset, or
  abort; declining/cancelling marks the run stopped with no fan-out. The gate
  fails closed if the dialog errors. Non-interactive hosts (env probes, CI,
  headless), resume replays, and the GHCP_MAESTRO_AUTO_APPROVE=1 bypass
  approve every subtask automatically, so existing paths keep working. The
  dialog uses stable per-subtask index keys (with enumNames labels) so
  duplicate agent names can't collapse a selection. 12 unit tests
  (tests/plan-approval.test.mjs).
• M6 — quality helpers. New runtime/quality.mjs builds four multi-agent
  patterns on top of spawnAll, each decoupled from any specific adapter and
  fully unit-tested with scripted adapters:
  • adversarialReview(findings, { reviewers, threshold }) — independent
    reviewers try to rebut each finding; survivors are those judged valid by at
    least threshold of their reviewers.
  • multiAngle(task, { angles }) — drafts the task from several angles in
    parallel, then a judge agent picks or merges the strongest draft.
  • fixLoop({ check, applyFix, maxIters }) — re-runs check until it passes
    (or maxIters), dispatching a fix agent between failed checks.
  • crossCheck(claims, { sources }) — verifies each claim across multiple
    perspectives and aggregates a support rate.
• M5 — saved workflows. New runtime/saved-workflows.mjs discovers,
  validates, and runs project/user/bundled workflow scripts:
  • scan order project (./.ghcp-maestro/workflows or
    $GHCP_MAESTRO_WORKFLOWS_DIR) > user (<dataDir>/workflows) > bundled
    (saved-workflows/), with kebab-case + reserved-name validation and
    priority-based de-duplication;
  • buildWorkflowApi injects a sandboxed api (bound spawn/spawnAll,
    phase, log, structured args, plus the M6 quality helpers) so scripts
    never touch the filesystem, shell, or SDK directly;
  • new commands /maestro run <name> [json|text args] and
    /maestro workflows; runs persist as workflow=saved:<name> so they show
    in /maestros and can be resumed;
  • bundled example saved-workflows/deep-review.mjs.
• Dev tooling + CI. Root package.json with test / lint / check
  scripts and an ESLint flat config (eslint.config.mjs). The runtime stays
  zero-dependency; ESLint and friends are devDependencies only. A
  no-console rule enforces JSON-RPC-safe logging in extension/runtime code.
• GitHub Actions: ci.yml (ESLint static analysis, node --check on every
  tracked .mjs, and the node:test suite across Node 20 and 22) and
  codeql.yml (CodeQL security-and-quality analysis).
• 43 new unit tests (tests/quality.test.mjs, tests/saved-workflows.test.mjs,
  plus buildPlanPrompt / sanitizeAgentName coverage), and 12 more for the
  M4.x approval gate (tests/plan-approval.test.mjs), for a total of 78.

Changed

• Extension load banner now reads … (M6 release) … and reports discovered
  saved workflows.
• Plan generation/parsing extracted from extension.mjs into the importable,
  pure runtime/plan.mjs (buildPlanPrompt, parseAndValidatePlan,
  sanitizeAgentName). tests/plan-parse.test.mjs now imports the module
  directly instead of regex-extracting the function source.
• Resume resolves both built-in and saved:<name> workflows via a single
  resolveWorkflowHandler.
• Line endings normalized to LF across the tree (per .gitattributes); the
  lockfile is now committed so CI can npm ci reproducibly.

Fixed

• Review hardening (CodeRabbit / CodeQL on PR #1).
  • quality.mjs: adversarialReview and crossCheck now group agent results
    by an exact per-item spec-id set (folding in the item index) instead of a
    string prefix, so findings/claims whose ids share a prefix (e.g. a and
    a-r1) can no longer cross-contaminate scores; crossCheck maps each
    verdict back to its source explicitly rather than by positional index.
  • quality.mjs: defaultSupportParser checks negative phrasing first, so
    NOT SUPPORTED / UNSUPPORTED / SUPPORTED: NO are no longer misread as
    supported.
  • saved-workflows.mjs: an unreadable (non-ENOENT) workflow directory is
    now recorded as skipped and skipped instead of aborting the whole scan;
    parseWorkflowArgs rejects JSON arrays so structured args are always plain
    objects; buildWorkflowApi helper wrappers tolerate a missing extra arg.
  • extension.mjs: failed resume and a missing saved-workflow descriptor now
    transition the run to error instead of leaving it stuck as running;
    task/brainstorm runs fail (status error) when every fan-out agent
    fails or when synth fails, and log per-agent failures otherwise.
  • plan.mjs: agent-id truncation length is now the named MAX_AGENT_ID_LEN
    constant instead of a magic number.
  • ESLint bans direct process.stdout/stderr.write in extension/runtime code
    (only session.log() is JSON-RPC-safe); CI/CodeQL checkouts run with
    persist-credentials: false; plugin.json declares its extensions entry
    and repository metadata explicitly.

M4 (previously unreleased)

• M4 — meta-prompt task workflow. /maestro task <natural-language task>
  launches a 3-phase run (plan → explore[N] → synth). The plan agent
  uses an LLM to decompose the task into 3-6 independent subtasks emitted as a
  JSON spec array; the runtime fans them out through the
  standalone-client adapter so each subtask runs in its own isolated
  Copilot CLI session. The synth agent merges the outputs.
• parseAndValidatePlan — fence-stripping JSON recovery with trailing-comma
  tolerance, schema validation (3-6 entries, unique non-empty agent,
  non-empty prompt, agent name <=60 chars, prompt <=4000 chars), and a one
  retry pass that feeds the parser error back to the planner.
• /maestro help (and /maestro with no args) prints all subcommands with
  short summaries, sourced from a single data-driven registry.