FlowCrew 0.4.1

Changed

• Bounded run-directory disk growth. Each codex stage's codex_home — transient CLI cruft (~90 MB/stage, dominated by the .tmp/plugins git packs, ~99% of a run's on-disk size) — is now purged automatically when the stage succeeds, and kept on failure/abort/timeout for debugging (dropping only the already-checkpointed SQLite WAL/SHM). Previously it accumulated unboundedly, so ~/.fc/runs could grow to hundreds of GB; it now stays at result-file scale (~KB per run). Extends the existing WAL-only cleanup — same automatic-cruft-cleanup category, no new config. flowcrew clean --keep N remains the manual escape hatch for pruning old run residue.

FlowCrew 0.4.0

FlowCrew 0.4.0 hardens the autonomous loop so that honesty becomes a property of the engine, not of the operator's diligence. It internalizes verify-before-trust, makes honest negatives first-class, adds self-remediation, and is validated across 6 task domains and 2 objective kinds (research metric and engineering acceptance).

Highlights

• Confirm-gate (verify-before-trust). A brief-declared research.confirm command runs as an independent check before any shipped terminal; a candidate that beats the metric but fails confirmation is downgraded, not shipped. Generalizes the Reality-Gate mechanism to the ship path.
• Honest terminal taxonomy. Added a first-class incomplete status (budget-exhausted mid-search — distinct from a crash and from an honest ceiling); unified the terminal-status guard so honest negatives are never mislabeled failed or scored at baseline by the outer loop.
• Anti-smoke integrity floors. Brief-declared research.integrity.field_floors reject token/"smoke" rounds (e.g. real_train_iters >= 150) instead of banking them as results.
• Self-remediation. Planner failures (empty/invalid dispatch.yaml) now bounded-retry with a specific re-prompt instead of a fatal punt (default_plan_stage_retries); the supervisor gains a REJECT verdict to send an insufficient deliverable back for re-work (default_supervisor_max_rejects); a deterministic no-progress stage watchdog aborts stalled stages.
• Fail-loud on degraded config. A workflow=research run whose frontmatter fails to parse now fails loud (a research_mode_degraded event) instead of silently degrading to plain dispatch.
• Observability precision. Integrity-rejected round counts surface in the terminal report; run summaries derive their decision label from the true terminal status; a declared confirm-gate that is skipped on a non-ship terminal is recorded rather than silently dropped.
• Literature scout + portfolio coverage. The outer loop can expand its portfolio from web search before exploring, with a deterministic coverage floor and domain-ledger awareness to avoid re-proposing dead directions.
• Regression-guarded. The hardening above is developed against a task-agnostic engine-contract suite (policy, terminal/verdict, dispatch retry/reject, stall detection, confirm-gate).

FlowCrew 0.3.0

v0.3.0 — Atom Architecture

flow-crew becomes a task-agnostic substrate: self-describing atomic semantics (roles, skills, checks, policies, terminal/verdict vocabularies) are collected in registries and injected into the planner, which composes a run and maps each semantic to the roles that execute it.

📖 Design & rationale: design/atom-architecture.md — the core idea, the drift problem it solves, the unifying abstraction, and the P0–P4 roadmap.

Highlights

• Engine fully task-agnostic — domain-specific logic lives in the brief / project contract; src/ carries zero domain vocabulary.
• Self-describing checks & skills + single-source vocabularies — no prose/enum drift between the engine and the planner prompt.
• Project acceptance contract (<project>/.flowcrew/contract.yaml) — the planner wires the contract's hard constraints as deterministic reality checks, written to reality_checks.md and enforced at terminal.
• Supervisor verdict parsing extracted into a pure parseSupervisorVerdict.
• Single-source descriptors for research policies, supervisor verdicts, and campaign-trigger types.

Research efficiency

• Significance margin — an improvement must beat the running best by more than ~1 standard error, filtering within-noise knob-tuning, plus explicit round-boundary steering toward genuinely different mechanisms.
• Single-source result schema (research.result_schema) — the engine validates each round's result against it and injects the schema into the planner.
• Reality-gate exec-script runs the script body through the shell from the project directory, fixing false-blocks on inline scripts and relative paths.

Informative dashboards

• Phases now show each attempt's research direction → result · duration. The campaign list stays cheap; only the detail view reads per-run journals.
• Knowledge digest — a campaign's per-run graphs synthesize into a ranked, deduped digest: best direction + result, findings, insights, dead ends, and cited sources, each linking back to the run that produced it. The relational graph stays available per run as a collapsible view, with draggable nodes.

FlowCrew 0.2.0

[0.2.0] - 2026-05-28

FlowCrew 0.2.0 is the first release positioned as a full agent workflow control plane rather than a browser-first alpha. It adds a Codex-default execution path, campaign memory, run knowledge graphs, deterministic Reality-Gate checks, supervisor steering, daemon/background task surfaces, and a rebuilt dashboard/documentation experience.

Highlights

• Codex-default execution: project defaults now route agent stages through Codex by default, while still supporting Claude as a per-run or per-role override.
• Recommended Claude Code -> Codex workflow: README and skills now present the intended loop: discuss and shape the plan in Claude Code, then /ship the confirmed task to FlowCrew for Codex-backed execution.
• Workspace dashboard redesign: added workspace/campaign views, run detail pages, campaign KPIs, phase progress, run lists, brief revision panels, and knowledge graph panels.
• Run Memory Graph: added dashboard and documentation support for persisted goals, approaches, findings, insights, results, user hints, and dead ends.
• Campaign intelligence: added campaign iteration state, brief versioning, diagnosis/review surfaces, cross-campaign knowledge graph support, plateau/regression triggers, and research workflow policy support.
• Reality-Gate framework: added deterministic evidence checks that can block unsupported terminal success, plus an audit command for retrospective verification.
• Supervisor and daemon runtime: added supervisor escalation, background task registry, daemon/task CLI surfaces, and run summary parsing.
• Release-facing documentation refresh: rewrote the README, refreshed screenshots and demo assets, and added focused reference docs for architecture, campaigns, CLI, configuration, Reality-Gate, and skills.

Changed

• Raised the package version to 0.2.0.
• Kept generated framework campaign logs local-only via .gitignore.
• Removed older dashboard screenshots and replaced them with current workspace, campaign, agent, architecture, and run-memory visuals.

FlowCrew v0.1.0

FlowCrew v0.1.0

This release turns FlowCrew into a more complete multi-agent execution system: task handoff, supervised execution, global run visibility, richer dashboards, and Claude Code-first defaults.

Highlights:

• Claude Code-first default: project defaults and CLI auto-detection prefer Claude Code while retaining Codex support.
• Global skill install by default: ./skills/install.sh installs /ship and /fc-status at user level, with --project available for repo-local installs.
• Supervisor brain: monitors running stages, produces progress briefs, and can guide, abort, re-plan, or detect completion.
• Global run workspace: stores runs under ~/.fc/runs so one dashboard can see all projects.
• CLI expansion: quick, status, list, guide, clean, export, doctor, start, and version commands.
• Dashboard overhaul: live output parsing, run summaries, task list caching, campaign rename, toasts, and monitor improvements.
• Agent handoff skills: shared /ship and /fc-status commands for Claude Code and Codex.
• Run memory and campaign intelligence: persisted evidence, decisions, campaign scoring, and pivot signals.
• Configuration and safety improvements: centralized defaults, supervisor config, path traversal checks, retry controls, and cleaner store paths.

Operational notes:

• The release tag points at commit e64711d.

FlowCrew v0.1.0-alpha

FlowCrew v0.1.0-alpha

Public alpha release for early feedback on the browser-first multi-agent workflow experience.

Highlights

• Browser-based multi-agent workflow runner for Codex
• Experimental Claude support
• Planner discussion before execution
• Visual plan review
• Live execution monitor
• QA gates and retry loops
• Persistent run history
• Campaign-aware task iteration
• Example workflows and demo assets

Install

npm install
npm run dev