Repository navigation
Releases
4.7.0: Evidence-backed preparation and on-demand capabilities
ForgeFlow 4.7.0 packages the capability foundation and review runtime for managed Claude Code and Codex installations.
- Choose among nine specialized procedures on demand after recording the task need, scope, available prerequisites and budget. Existing permissions and exclusions still apply.
- Preserve immutable source/input evidence with exact retrieval and freshness checks.
- Keep focused specialist assignments and bounded follow-up requests, with explicit unresolved outcomes.
- Prepare scope, packets, budgets and bounded review waves through one operation. Return ready inputs or named blockers while retaining required coverage, proof and safe recovery.
- Keep ordinary feature validation proportional; comparative studies require a separate request.
- Refresh the README, wiki and generated reference pages.
Validation: a fresh disposable checkout with Node 24.21.0 and locked dependencies passed 227/228 deterministic test commands initially. The sole failing assertion expected obsolete capability-boundary wording; after correction, the focused context-packet check passed. Final documentation/link/version/wiki checks, source smoke, strict bridge typecheck and public-summary rendering passed. Both managed hosts installed successfully from the runtime archive into disposable homes; installed bytes match the archive, and archive bytes match the release commit. Existing owned-PR functional evidence was reused. No model comparisons were run. GitHub Actions remains disabled.
Download forgeflow-4.7.0-runtime.tar.gz and SHA256SUMS, verify with sha256sum -c SHA256SUMS, extract the archive, then run node scripts/forgeflow/install-template.js --target codex or --target claude from its directory. Restart host sessions after updating. This archive includes managed runtime files and documentation; evaluation fixtures and completion study runners are excluded. Use the source checkout for research reproduction.
Limits: all nine procedures retain evaluation benefit labels. No general accuracy, speed or cost advantage is claimed. Live client restart/discovery, manual settings, additional operating systems, live providers, public deployment and physical fabrication remain separate qualification. The lean Codex plugin does not expose the managed capability entry point. GitHub's automatic source archives retain historical research material; no new redistribution-rights claim is made for upstream-derived fixtures.
4.6.3: Keep workflow context local
Keep internal workflow context out of application changes and public handoffs.
- Add consistent local-only rules across both supported hosts.
- Generate PR drafts without private evidence identifiers, descriptions, or report paths.
- Add local Git exclusions and block publication of tracked session state, including files present in outgoing commit history.
- Refresh the generated host bundles and packaged metadata.
Validation: Node 24 local suite completed with 198/202 commands passing initially. All four failing documentation and generation checks passed after corrections. Source smoke, Pi smoke, strict typecheck, public-summary rendering, generated prompt consistency, and whitespace checks passed. GitHub Actions remains disabled.
Limitations: workflow attribution requires inspection of code and public text; automated publication checks cover local state paths. Restart host sessions after updating to load the new instructions.
ForgeFlow 4.6.2: Complete local shipping previews
Shipping previews now include staged and unstaged edits and nonignored new files alongside committed branch changes. Previously, uncommitted work could be missing from the preview.
- Each path appears once, and the reported count uses the same inventory.
- Staged changes remain visible when an unstaged edit restores the base content.
- Filenames containing tabs or newlines are preserved. Renames appear as a deletion and an addition.
- Task validation, evidence freshness, and reviewer approval remain separate from the file inventory.
Validation: ran npm ci --ignore-scripts and the full npm test suite locally with Node 24.21.0. The initial run passed 200 of 202 test commands; the two documentation checks passed on rerun after correcting the wiki release link. Service tests, the strict bridge typecheck, shipping regressions for both installed host layouts, shell syntax, and whitespace checks passed. Source smoke, the Pi smoke check, release-to-install preflight, and the public-summary fixture rendering also passed.
The preview does not stage or commit files. GitHub Actions remains disabled.
ForgeFlow 4.6.1: Clearer user communication
ForgeFlow 4.6.1 improves how agents explain their work: match the reader's background, report useful findings during progress updates, describe concrete outcomes, and keep validation limits clear. It retains Orwell's six rules and preserves exact technical details and required output formats.
The update covers Claude agents and core commands, Codex workflow skills, and generated Codex agents. README and wiki guidance also clarify agent responsibilities, setup, migration, and recovery.
Validation: all 202 local test commands passed on Node 24.21.0, including service tests and strict typechecking. The committed source matches the tested worktree byte-for-byte. The initial test-recording attempt timed out at its one-minute limit; the complete rerun passed with a longer limit. Skill validation, agent prompt consistency, documentation links, and wiki export checks passed.
Two live agents produced four sampled responses covering a handoff, stakeholder JSON, and a hypothetical progress update. The samples preserved evidence limits and output formats; a retest reduced unnecessary internal detail. This is a bounded check, not a user study or measured productivity claim.
Existing 4.6.0 installations can use the current updater. Restart Codex and Claude Code after updating.
For the first upgrade from legacy agent names, follow the source-checkout migration procedure. Use FORGEFLOW_RELEASE_SHA=6a90c6aebca3c6a76855a76fa7e12502bb974a91. Do not begin that migration with an old installed updater.
Changelog · Communication guide
This is a source release; no separate compiled binary is required. Validation ran locally, and GitHub Actions remains disabled.
ForgeFlow 4.6.0: Clear agent roles and safer upgrades
Upgrade from 4.5.0 or earlier
Use the installer from this release checkout for the first upgrade from legacy agent names. Do not run the old installed /update-forgeflow or $update-forgeflow for this migration. An old updater can remove customized legacy agents before its new code takes effect. If that already happened, use its rollback before another install replaces the backup.
Release commit: 7fda3c5bbb75fd5111ec6ae950e87a7e31733708.
In Bash:
FORGEFLOW_MIGRATION_DIR=$(mktemp -d)
git clone https://github.com/BrandedTamarasu-glitch/ForgeFlow.git "$FORGEFLOW_MIGRATION_DIR/source" || exit 1
cd "$FORGEFLOW_MIGRATION_DIR/source" || exit 1
git checkout --detach 7fda3c5bbb75fd5111ec6ae950e87a7e31733708 || exit 1Then choose your host and preview before installing:
# Claude Code
node scripts/forgeflow/install-template.js --target claude --dry-run --json
node scripts/forgeflow/install-template.js --target claude
# Codex (honors CODEX_HOME)
node scripts/forgeflow/install-template.js --target codex --dry-run --json
node scripts/forgeflow/install-template.js --target codexUse only your host's pair, or both if you use both hosts. The installer preserves edited or unverified legacy agents and reports them as preserved_legacy. Merge those instructions deliberately into the new roles. A full install can replace customized current managed files, so keep an independent copy of changes you want to merge. Restart your host to load the new names.
Future upgrades use the new updater installed by this step. An unchanged reinstall preserves the recovery snapshot; a later changed install or repair replaces that single snapshot. See the upgrade and rollback guide for custom homes and recovery.
Changes
- Clear role names: Smith → Builder, Warden → Guardian, Lumen → Designer, Atlas → Coordinator, Arbiter → Architect, Compass → Product Lead, Aegis → Verifier.
- Explicit task context in CLI and dashboard messages, retained through replay and export. Historical evidence stays intact.
- Orwell's six writing rules in agent prompts for clean CLI output; specialized debate judging and answer-key validation instructions.
- Stable historical learning IDs, correct mixed-name report totals, and context rebuilds that preserve old packets without validating them as current output.
- Narrow automatic legacy cleanup, preserved no-op rollback snapshots, complete lifecycle message limits, missing-catalog hook recovery, and bounded profile registration.
- Current service contracts, schema examples, role guides and presentation labels.
Validation and limits
On release commit 7fda3c5, the full local Node 24 suite passed 202/202 test commands, including service tests and TypeScript. All 28 browser checks passed against an isolated dashboard. A fresh npm ci --ignore-scripts completed, and version, plugin-manifest and documentation-link checks passed. Runtime code matches the independently reviewed audit fixes; release-only changes synchronize version manifests and public documentation. GitHub Actions remains disabled.
No live host restart, hosted model run, or measured model-performance improvement is claimed. The first-upgrade route above is required; publishing this release cannot change an already-running old updater.
ForgeFlow 4.5.0
Changes
Shipping summaries now use the explicitly selected task's current evidence. The helper checks source and artifact freshness, uses each criterion's latest result, and keeps automated tests, manual checks and review evidence separate. Historical notes no longer become passing tests or approval; missing, stale and failed proof remains visible.
Also included since 4.4.0:
- Optional Obsidian vault memory and portable handoffs.
- Task-focused advice and memory selection, with proportional design and validation guidance.
- Checks for workaround durability, tracking unrelated bugs, user impact and training, and actionable errors.
- Optional RTK installation, safer Codex generator option handling, and Ember/dashboard startup fixes.
- Local validation for the ForgeFlow repository, with GitHub Actions disabled and GitHub reserved for release pushes and artifacts.
Using the change
Run ship-prepare.sh --task <id> "Title" for the task matching the current objective. Existing invocations without a task still prepare artifacts but report validation as missing. Curate relevant historical notes manually, and verify explicit reviewer verdicts against the current source before publication. Task readiness is not review approval.
Validation
On release commit a0a8a85, the complete local runner passed 199/199 test commands using Node 24.21.0, including service tests, TypeScript checks and installed-host regressions. All eight source smoke checks passed. Four specialist agent reviews approved; Arbiter approved and Compass confirmed. All 18 reviewed file hashes match the committed source.
No GitHub-hosted jobs are required. Task evidence covers the task's declared scope; this release does not claim automated enforcement of agent compliance or model performance gains.
ForgeFlow 4.4.0: Task evidence and recovery
ForgeFlow 4.4.0 connects a task objective to acceptance criteria, current-source proof, and resumable phase history in Claude Code and Codex.
New capabilities
- Start and inspect tasks with
/taskin Claude Code or$taskin Codex. Checks bind results to source and artifact hashes; changed evidence becomes stale while history remains available. - Inspect task criteria, proof, phase history, and next action in the workshop. Background verification keeps the dashboard responsive.
- Plan balanced no-agent, single-agent, and ForgeFlow comparisons with explicit actual, fixture, and unknown outcomes.
- Suppress stale or contradicted memory through dependency evidence and task-linked feedback.
- Validate fleet ownership and resource declarations; turn supplied CI failures into bounded local repair/check/draft workflows.
Reliability
Oversized task writes preserve the last readable record. Fleet checks committed, staged, unstaged, and untracked changes independently. Task scans share verification within each scan, bound concurrent workers, and deliver timeout errors promptly while cleanup completes.
Also included since 4.3.72
The refreshed workshop and Ember companion, session dashboard opening, visual HTML/PDF guide, rewritten wiki onboarding, Codex model inheritance, research divergence tools, memory and command-learning controls, runtime hardening, and deterministic CI.
Update
Use /update-forgeflow in Claude Code or $update-forgeflow in Codex. Restart the host and any running dashboard to load the new runtime. Existing workflows without task records remain supported.
Task evidence requires a Git worktree with an initial commit. Recovery restores saved phase context, not hidden model state. Initialized Git submodules remain unsupported by whole-repository snapshots. Fleet checks declarations without reserving infrastructure; maintenance prepares local drafts without publishing a PR. Evaluation fixtures do not establish model superiority or user adoption.
Validation
195 test commands and 25 browser tests passed for the implementation and corrective review. A fresh dependency install and 195/195 test commands also passed in a clean release checkout. GitHub Actions passed on the exact release commit, including the clean-tree check.
Forgeflow 4.3.72 - Canonical Codex Agent Parity
Forgeflow 4.3.72 synchronizes Codex execution with the canonical Claude agent contracts.
Highlights:
- Regenerates all 28 Codex agents from full canonical prompts with SHA-256 drift checks.
- Adds Atlas and Compass presentation agents.
- Installs Codex runtime helpers, templates, patterns, and agent-chat resources under CODEX_HOME/forgeflow.
- Adds installed-runtime fallback guidance to Codex skills.
- Separates Codex-native skill support from generic host wrappers in capability reporting.
- Avoids the Codex built-in /review collision.
- Makes freshness-sensitive release tests deterministic.
Validation:
- Full release-readiness gate: ready, 146 checks, zero blockers.
- Source smoke: pass.
- Codex agent drift: 28 agents passed.
- Codex agent TOML parsing: pass.
- Install-template, plugin, docs, command coverage, and whitespace checks: pass.
See docs/changelogs/v4.3.72.html for full notes.
Forgeflow 4.3.71
Forgeflow 4.3.71 improves lean review precision. /forgeflow-lean-review findings now include confidence, replacement guidance, estimated net lines, why-safe and why-not-safe evidence, and proof steps. The helper also suppresses obvious semantic false positives into visible skipped-boundary records instead of producing noisy findings.
Docs updated: README, hosted docs, wiki Home, Workflow Commands, and Context Intelligence.
Validation: node scripts/forgeflow/test-render-lean-review.js; node scripts/forgeflow/test-release-version.js; node scripts/forgeflow/test-doc-links.js; node scripts/forgeflow/test-command-coverage.js; node scripts/forgeflow/test-install-manifest.js; node scripts/forgeflow/test-update-forgeflow.js; node scripts/forgeflow/test-health-check.js; node scripts/forgeflow/test-runtime-helper-contract.js; node scripts/forgeflow/smoke-check.js --mode source --json; git diff --check.
Forgeflow v4.3.70
Forgeflow 4.3.70 wires lean guidance into the normal review and ship path. /review now attempts the lean review helper automatically when available, and /ship plus ship-prepare record lean-readiness evidence for lean-path work while preserving the existing review, secret-scan, validation, PR, push, and CI gates.
Validation: release-version, doc-links, command coverage, release notes draft, install manifest, update-forgeflow, runtime helper contract, health check, and source smoke all passed.