FailSafe: Governance for AI Coding Agents
What's New in v6.0.13
- Statewright AI in the Integrations tab. A Statewright card opens Statewright's own Run History website (
https://statewright.ai/runs) in your browser when you choose it. Statewright handles sign-in and run data; FailSafe sends nothing to Statewright, shows no connection status, and does not read Statewright history or enforce Statewright workflows (#652).
- Witnessed release ledger. The governance ledger behind this release is checkpointed with an owner-held, non-exportable TPM key and an independent RFC 3161 timestamp, and the checkpoint is verified before publication (#638). This is release tooling; it adds nothing to the installed extension.
- Continue and Aider stay disabled (#475). See the changelog for known limitations.
What's New in v6.0.12
Released 2026-11-10. One theme for the whole app: the theme you pick in the Console now also applies to the Monitor sidebar and the SRE view, stays the same in the browser Console and in VS Code, and is kept across reloads and restarts in the VS Code setting failsafe.console.theme (#601). Includes the dependency updates merged after v6.0.11.
- Continue and Aider no longer start from FailSafe. FailSafe cannot yet contain them at the operating-system level, and a git worktree is not a sandbox, so both commands refuse before any process, prompt or credential read, and the catalog shows them disabled (#475). This is a temporary restriction.
- Qor-logic is pinned to the frozen 0.176.1 release. FailSafe installs it only from the verified local wheel you select, offline, with no PyPI upgrade, and reports a skill install as successful only when the CLI's file count matches a hash-verified install receipt. Selected-skill and global installs, the install preview and the skill list are unavailable with this release (#463). See the changelog for known limitations.
Understand agent activity, check risky changes, and keep evidence of what happened.
FailSafe runs locally in VS Code and Cursor. It monitors agent and workspace activity, evaluates deterministic policies, and presents risks, decisions, and checkpoints in one Console. Some controls can stop a governed action; editor-save denials are reported but cannot prevent VS Code from saving a file.
What FailSafe Does
- Understand changes and risks. Monitor file activity, inspect Sentinel findings, and review workspace risk and governance state.
- Establish intent before execution. Associate work with a defined intent and inspect policy decisions before approving higher-risk actions.
- Review agent changes. Govern supported agent workflows and inspect changes before applying them. Some integrations observe activity rather than prevent all writes.
- Preserve an audit trail. Record decisions, checkpoints, and provenance so operators can investigate what happened and what remains unverified.
- Navigate architecture and delivery evidence. Workspace Architecture and Delivery connect human-reviewed architecture and implementation evidence.
Where to start: Open the FailSafe Console for workspace status and risks. Use the Governance views for decisions, and the Audit Log for evidence. Full feature reference.
Quick Start
- Install FailSafe from the VS Code Marketplace or Open VSX.
- Open a workspace in VS Code or Cursor.
- Run
FailSafe: Open Console (Browser Popout) or press Ctrl+Alt+F.
- Run
FailSafe: Set Up Agent Governance to inject governance rules into detected agents.
- Run
FailSafe: Audit Current File to generate a governance verdict for the active editor.
- If you want commit-time guardrails, run
FailSafe: Install Commit Hook inside a git workspace with curl available on the path.
What's New in v6.0.10
- The Console shows your workspace's text as text. Ledger verdicts, plans, risks, skills, alerts and integration fields are escaped everywhere the Console shows them, so crafted workspace content can no longer run script in the Console page.
- Other web pages can no longer reach the Console's dialogs or live updates. Live updates accept only the Console's own page, and the Console answers only requests addressed to a loopback name. Another page cannot open the L3 approval or Panic Stop dialogs, or resume the Sentinel or run Verify Integrity. Still open: a few body-less actions (B384),
show-output (B383), and re-checking the chain by loading the checkpoint list (B396), and running git repeatedly by loading the tracker data (B402).
- Dependency security. The MCP server refuses a tool call with more than 10,000 argument elements, the Sentinel file watcher moves to chokidar 4, and the production dependency audit reports no findings. The chokidar 4 watcher change was not tested on a Mac (B368).
- Enforce mode works again in the VS Code extension host. You can approve or veto, and then seal, the active intent, and forged intent files are not honoured for editor writes or governance verdicts. A denied editor save is still reported, not prevented. A workspace writer can still switch enforce off (B371), the ACP proxy is not covered (B373), some displays still read the intent file (B370), and the release gate and the commit check still read it (B374).
- Workspace. Record a delivery review, including a declared rework of a vetoed commit; see contradicted or missing design authority in Architecture; and compare Taxonomy with the Console's sub-views, with save containment (B330); on Windows a folder swapped during a save is narrowed, not fully closed (B332).
- Behaviour change. A Console opened through Codespaces, vscode.dev or code-server behind a proxy name is refused (403) on every page and live update. Use a local browser or the VS Code commands.
See the changelog for compatibility changes and known limitations.
Optional Integrations
FailSafe connects to the tools your AI agents use. Every integration is opt-in. Integrations that call an outside service stay off until you enable them in settings and store a credential with FailSafe: Set Integration Credential; the installers and the risk scorer do nothing until you run them.
A recent call is not end-to-end qualification or proof of customer value. An integration status describes the latest observed attempt, not whether every promised action works. Configured is not connected, connected is not fully qualified, and an AI agent's unreviewed changes are not proven safe.
FailSafe's internal governance MCP service currently runs on an in-process stdio transport and is not reachable by external MCP clients. Supported integration clients are not a universal tool interceptor. FailSafe does not ship an Omium outcome-verification adapter; that proposed bridge is separate from this release.
Supported agent integrations can add controlled workflows and evidence:
| Integration |
What it does |
Why it matters |
🤖 Continue (cn) governed command (disabled) |
FailSafe: Run Continue (governed) refuses before launching Continue. |
A disposable git worktree is not an OS sandbox, so FailSafe does not claim agent containment it cannot enforce. No Continue process starts and no setting or credential is read through this command (#475). Historical Windows CLI limitation: Continue's npm shim cn.cmd cannot run without a shell; this does not override the current no-launch gate. |
| 🔧 Aider governed command (disabled) |
FailSafe: Run Aider (governed) refuses before launching Aider. |
No Aider process starts and no credential is read through this command until a separately qualified containment mechanism exists (#475). |
| 👁️ OpenHands run import |
Import an exported OpenHands run, after you confirm, as historical evidence in the FailSafe timeline (FailSafe: Import OpenHands Run (observe)). Version-gated, deduplicated and signed for this install; the transparency log gets one summary line. |
See what an agent run did next to your own governance events. It is a record of a past run, not live enforcement (#477). |
| 🔎 Cline / Roo / Kilo policy audit |
Scan workspace MCP/tool config and flag risky posture: remote MCP, wildcard auto-approval, shell-capable tools (FailSafe: Audit Agent MCP Policy (Cline/Roo/Kilo)). |
Catch an over-permissioned agent before it bites; secrets in the config are redacted before any finding is recorded. |
Connect your issue tracker, security, and team tooling:
| Integration |
What it does |
Why it matters |
| 📥 Linear / Jira intent drafts |
Fetch one Linear or Jira issue, review it in an unsaved preview, and create a governed intent draft only when you approve (FailSafe: Import Linear Issue as Intent Draft, FailSafe: Import Jira Issue as Intent Draft). Read-only toward the tracker. |
The ticket becomes reviewed intent without retyping it. Nothing is written to Linear or Jira and nothing is synced (#480). |
| ✅ GitHub advisory checks |
Publish a neutral, manual advisory Check Run on the current commit (FailSafe: Publish Manual GitHub Advisory). When the commit has a pull request, the same command also posts a PR-linkage Check Run. It is not a SHIELD verdict. |
The SHIELD check name is reserved for a verified, commit-bound receipt. No command publishes a verified SHIELD check yet. Fork PRs degrade to local-only. |
| 🐞 Sentry regression import |
Pull a Sentry project's unresolved issues into the risk register as runtime-regression risks (FailSafe: Import Sentry Regressions), with first and last release from each issue's detail. |
Production failures become governed risk records tied to project / environment / release, with an exact count of what was created, updated, unchanged or failed; your triage survives re-imports. No raw event payloads, and no commit correlation. |
| 🛡️ SARIF security ingestion |
Import Semgrep / CodeQL / any SARIF 2.1.0 scanner output into the risk register (FailSafe: Import SARIF Findings). |
Your security scanner stops being a separate silo: every finding becomes a governed risk in the same audit trail as agent decisions. |
| 📣 Slack / Microsoft Teams notifications |
Post VETO / L3-approval / drift events to a Slack or Teams webhook. Notify-only, off by default. |
Governance becomes a team signal: the right people see a blocked action or a queued approval in their channel. |
| 🧮 MCP Registry risk scoring |
Score any MCP server locally: read-only, with field sanitization: before you trust it. |
Adopt MCP servers on evidence, not vibes: a supply-chain admission check at the door. |
| 📦 MCP Catalog installers |
Risk-scored installs of Context7, Mermaid Chart, and Playwright MCP into your .mcp.json, written only after you confirm (FailSafe: Install MCP Integration (governed)). |
Governed installs of tools that make your agents measurably better, with the trust check built in. |
| 🧠 Bicameral MCP |
Detect, connect, and ratify architecture decision records and their drift inline. |
Every Bicameral tool call passes through FailSafe's universal interceptor. |
| 🎨 Open Design |
Observe Open Design agent runs and read its projects through a read-only MCP connection. |
Artifact creation (create_artifact) is refused in this release, even after an L3 approval, until approvals can be verified where the call is made (#481). |
| 🧰 Agent Governance Toolkit installer |
Detect your workspace environment and pre-fill the matching AGT install command in a terminal; you run it. |
One entry point to the right per-language or per-host package. |
Each integration ships with its own README and its external API names are back-cited to official docs. A Tier 1 supply-chain CI baseline (least-privilege tokens, SHA-pinned Actions, dependency review, CODEOWNERS) hardens the repository itself.
Agent Marketplace: install, update and scan are disabled until isolated setup and fail-closed scanning land (#468, #469).
Open the Integrations tab to configure an integration. Each card shows the result of that integration's last real call (Verified, Connected, Failing with a reason, or Stale after 7 days); an integration that is only configured shows Configured · unverified.
Current Release: v6.0.13 (2026-11-13)

What's New in v6.0.7
- The ledger and Shadow Genome now actually store data on Windows and macOS. v6.0.6 shipped one package with a Linux-only SQLite binary, so both ran in stub mode everywhere else. FailSafe now ships a separate package for each platform (Windows, Linux, Alpine and macOS, each on x64 and arm64), each carrying only its own binary. VS Code 1.101 or later is required.
- Decisions you make from the Console are confirmed in VS Code. Approving or rejecting an L3 item, ratifying a Bicameral decision, rolling back to a checkpoint, panic stop and turning off the agent hooks now ask you to approve a VS Code dialog that shows what will be recorded. Before, any program on your machine could make those calls.
- Continue and Aider run in a throwaway copy of your repository. Their changes wait for your review and reach your workspace only when you choose Apply. Child processes the agent leaves behind are found and stopped before the result is checked.
- Deleting or editing a risk no longer wipes out imported risks. The risk register used to rewrite the file from a copy taken when VS Code started.
- Editor-save denials say what they are. VS Code cannot cancel a save, so FailSafe now reports a denied save as observed, not prevented, instead of claiming it was blocked.
- The SHIELD phase reads correctly on Windows checkouts, where Git stores the ledger with CRLF line endings.
- Sentry imports report exactly what happened and respect Sentry's rate limits; imported risks show the issue's first and last release.
- Integration credentials move to Secure Storage. Tokens, API keys and webhook URLs for Sentry, Jira, Linear, GitHub, Continue, Slack and Teams are no longer read from settings. Use FailSafe: Set Integration Credential; if FailSafe finds one in plain-text settings it offers to move it, checks the stored copy, and removes the plain text only with your consent.
- The L3 Approval Queue buttons work. Approve, Reject and View File were blocked by the panel's own security policy.
- An interrupted save no longer corrupts the risk register.
risks.json is replaced only by a complete file.
- Two VS Code windows can no longer overwrite each other's risk changes. Every update to
risks.json takes a short lock and re-reads the file first, and re-imported findings keep the triage status you set.
- The SOA ledger notices when it is rolled back or swapped. Every column of a new row is sealed, and the ledger's head is anchored in VS Code Secure Storage, so emptying, truncating or replacing the ledger is reported.
- The SHIELD check name on GitHub is reserved for a verified, commit-bound receipt. A receipt is verified against the governance ledger in Git history. No command publishes a verified SHIELD check yet. FailSafe: Publish Manual GitHub Advisory posts a neutral advisory.
- FailSafe runs Git only from a checked location. On Windows, a
git.exe placed in an opened folder could run instead of Git; FailSafe now uses a verified Git outside your workspace and never runs Git through a shell.
- Open Design 0.24.1 works again, and the Open Design client can no longer approve its own artifact writes.
Not verified against live services in this release: Sentry, Jira, Linear, OpenHands, Aider and Continue were tested with recorded responses only. See the [6.0.7] changelog for the full list of known limitations.
What's New in v6.0.6
- Screen readers now hear why a Shadow Genome node failed, not just that it did: failure nodes carry their severity in the accessible name.
- A status row that lied when it couldn't reach upstream. The Bicameral panel reported "0 open issues" whether the upstream was empty or unreachable. Those now look different.
- The Audit Log records which engine verified an entry, so an existence check is no longer indistinguishable from a content check.
- A voice glitch that blocked unrelated updates is gone. A timing check measured itself against a stopwatch that could round the wrong way, so it failed at random and held up routine dependency updates.
- Generated test screenshots stop landing in the repository. An ignore rule pointed at a folder path that does not exist, so it never took effect.
- The system-state document tells the truth again. It reported v5.9.0 as current while v6.0.4 was shipping, and nothing checked it. Something does now.
- The governance index distinguishes "deliberately private" from "missing": and one entry turned out to point at a file that has never existed in this repository.
- Security:
qs and fast-uri dependency updates.
Under the hood this is a governance-integrity release. Five checks were found reporting success while inspecting nothing at all; each now has a test that proves it can fail. Four issues were filed upstream against Qor-logic.
What's New in v6.0.4
- "Install / Refresh Skills" now actually upgrades a stale install. The button promised to bring qor-logic up to the required version, but it only checked whether the package was present - so an install sitting below the floor was treated as fine, and the next step ran against the outdated CLI while reporting success.
- Commit checks are faster on large repositories. The workspace snapshot behind the pre-commit guard read and parsed the governance ledger five times per build - 8.7MB and 477ms cold, on the path that blocks your commit. It now reads once.
- Development Tracker phase chips are readable by screen readers. Each chip's status is announced rather than conveyed by colour alone.
- Audit Log records commit to the content they judged. A warning or block now carries a hash of the exact content the verdict was made against, so a later decision can prove it refers to the same bytes.
- The governance ledger cannot silently fork. A new pre-merge check fails CI when concurrent work would write conflicting ledger entries.
What's New in v6.0.3
- The Observe-to-Enforce upgrade notice now actually appears. v6.0.0 changed the governance-mode default and shipped a one-time notice to tell existing installs their behavior had changed. That notice could never fire, so every install predating 6.0.0 that had never explicitly picked a mode moved to Enforce silently. Fixed, and the detection now uses the same technique first-run onboarding already relied on.
- The Development Tracker stays responsive on large repositories. Its PR-cadence read ran an unbounded
git log on every dashboard load: over a second and past the output buffer on a 150k-commit repository, where it failed silently. It is now bounded, and says so when the window truncates instead of degrading invisibly.
- Voice synthesis cannot wedge the Mind Map. A stalled speech-synthesis call is now bounded and surfaces as an error rather than leaving the voice controls stuck.
- Mind Map tells you how dense the graph is. A live
N nodes · N edges readout, including when duplicate edges were merged.
- The ACP proxy notices if its registry entry is rewritten. That registry is user-writable, so an outside change could quietly route around governance while still showing the governed name. Activation now distinguishes intact, tampered, missing, and malformed registries: and never blocks activation if the check itself fails.
- Audit Log entries link to their resolution. A warning or block now records the exact ledger row a later decision resolves, by id rather than by guessing from path and timestamp.
What's New in v6.0.2
- The Monitor works without a mouse. The sentinel warning banner, critical-blockers graphic, error-budget readout, the five workspace-health metric cards, and governance-alert rows are all keyboard-reachable with visible focus rings and screen-reader names; Enter opens the explanation or details modal, and Escape returns focus to where you were: even when a live refresh rebuilt the list underneath the modal.
- Audit Log records now say what, where, and why. A sentinel verdict used to render as bare identity JSON; records now name the triggering file, the human summary, and which heuristics fired. A severity chip row: All levels · Issues · Warn · Block: isolates what needs attention, and the CSV export is a real RFC-4180 CSV with the columns you triage by (it was previously unparseable).
- Corrupt files are preserved instead of destroyed. If the risk register, adapter config, or marketplace state exists but is unreadable, the original is set aside as a
.corrupt-<timestamp>.bak before anything overwrites it. Previously the first write replaced the file and the data was gone.
- The risk register no longer absorbs your backlog. Adding or importing a risk on a fresh workspace used to durably persist the entire
BACKLOG.md projection as if you had authored every row; mutations now touch only the durable store, and backlog-derived rows are shown read-only.
- Sharper governance signals. The PR-linkage check reports accurate reasons (a pull-request number is named as one, and unverifiable references warn instead of falsely failing), the governance webhook's payload is provably what was signed and fails closed when oversized, credential redaction covers passwords containing
@, and sidebar Monitor modals draw their proper overlay chrome.
What's New in v6.0.1
- The Monitor's warning banner is now a working link. Clicking "N issue(s) detected" opens the Console's governance Audit Log focused on the triggering verdict: from the embedded sidebar (a webview-safe relay replaces the sandboxed
window.open, which used to silently do nothing) and from browser-served Monitors alike. Verdict events carry their own timestamp through the checkpoint and transparency logs so the deep link lands on the exact record, and verdict deep links bypass the Audit Log's default same-day date filter. The blockers and error-budget click-throughs use the same relay.
What's New in v6.0.0
- The pre-commit guard is real. The commit-check endpoint ships (it was documented but never implemented), the hook's port tracks the live Console server, and installing from a git worktree governs every worktree of the repo.
- Guided Agents-window setup.
FailSafe: Configure VS Code Agents Window Governance walks the opt-in, worktree commit-hook install, and governed MCP integration installs.
- Repository-scoped first-run. Each repository gets its own governance-mode decision; an explicitly configured mode anywhere suppresses re-prompting.
- Accessible Mind Map. The graph carries an accessible name and a LIST VIEW toggle renders real node/edge tables: same pattern as the Shadow Genome table view.
- Integration hardening + a 69% smaller package. MCP installs can no longer destroy an unparseable
.mcp.json; SARIF/Sentry imports are malformed-proof and never silently drop findings; the PR-linkage check paginates (no more false findings); build intermediates and governance scan outputs can never ship in the VSIX.
- Enforce is now the default governance mode. New and never-configured installs gate writes out of the box (intent-gated saves, L3 approvals). Observe and Assist are unchanged and one command away (
FailSafe: Set Governance Mode); a one-time notice on upgrade offers the mode picker. Unresolvable mode values and a missing ACP mode mirror now fail closed to enforce.
- Enforcement works on every tier. The editor enforcement path no longer consults any feature gate.
- Enforce-mode Create Intent flow. Creating an intent in enforce mode first selects the plan it serves, with an explicit switch-mode escape; the writes-blocked dialog offers
Set Governance Mode.
- Cleaner product surface. The observe-mode advisory banner and settings hint are removed, and
/api/v1/status now reports the true governance mode.
- Agent Skills marketplace category. Install the MIT-licensed mattpocock/skills packs (Wayfinder decision-ticket planning + engineering/productivity companions) as governed, risk-scored marketplace entries.
- Mind Map view prefs survive reload on any machine. Fixed a view-preferences identity race caught by the release gate.
- More fail-closed hardening + accessibility. GovernanceRouter verdict faults now block instead of silently allowing; malformed verdict payloads escalate instead of silently PASSing; a malformed META_LEDGER reads as damaged, not idle; L3 escalation-queue failures fail visibly; ACP mirror-write and fs no-client paths fail closed; real ARIA tab semantics on the Command Center nav and sub-view pills; Space push-to-talk guarded against focused controls.
What's New in v5.9.0
- 📌 The Development Tracker stays put: no more reloads when live data refreshes, it fills the available space, and you can export it as a clean PDF.
- 🌱 The Mind Map starts from your repo: an empty map preloads a knowledge graph projected from your governance history, kept distinct from your own brainstorm work.
- 🗂️ Workspace › Taxonomy editor: review and edit the tracker's programs, verticals, and agent mappings; Save writes a governed
tracker-config.yaml the assistant is directed to consult on its next cycle.
- 🧬 Shadow Genome tells the truth: the summary card now reads "Graph Nodes" (not "Failure Nodes"), "Observed" can no longer contradict the failure count, and dense graphs open in a readable table.
Previous releases
The full release-by-release history (v5.x and earlier) lives in CHANGELOG.md. Skills are sourced from the qor-logic PyPI package.
Bundled Documentation
docs/COMPONENT_HELP.md - every shipped surface, metric group, and governance component in one place
docs/PROCESS_GUIDE.md - setup, audits, commit hooks, provenance, break-glass, replay, rollback, and troubleshooting flows
Core Commands
| Command |
Purpose |
FailSafe: Open Console (Browser Popout) |
Open the main governance console in a browser window |
FailSafe: Open Console (Editor Tab) |
Open the governance console in an editor tab |
FailSafe: Audit Current File |
Run a manual audit on the active file |
FailSafe: Set Governance Mode |
Switch between observe, assist, and enforce |
FailSafe: Set Up Agent Governance |
Detect supported agents and inject governance rules |
FailSafe: Install Commit Hook |
Add the authenticated pre-commit governance hook to the current repository |
FailSafe: Remove Commit Hook |
Remove the FailSafe pre-commit governance hook and token file |
FailSafe: Activate Break-Glass Override |
Start a time-limited emergency override |
FailSafe: Replay Verdict (Audit) |
Re-run a prior governance decision for comparison |
FailSafe: Revert to Checkpoint (Time-Travel) |
Restore a recorded governance checkpoint |
QorLogic: The Governance Layer
QorLogic is the deterministic governance engine that enforces safety policies at the editor boundary. It operates on a fundamental principle: governance decisions are made by code, not by asking an LLM to follow rules.
Prompt Guidelines vs. Deterministic Governance
| Aspect |
Prompt-Based Safety |
QorLogic Deterministic Governance |
| Decision Maker |
LLM interprets rules |
TypeScript code executes rules |
| Consistency |
Varies with context, temperature, model |
Identical output for identical input |
| Auditability |
Opaque reasoning chain |
Explicit code path, logged decisions |
| Bypass Risk |
LLM can ignore or reinterpret |
Code cannot be persuaded |
| Speed |
Network latency + inference |
Sub-millisecond local execution |
How QorLogic Works
Risk Classification: Files are classified as L1 (low), L2 (medium), or L3 (high) risk based on:
- File path triggers (e.g.,
auth/, payment/, credential → L3)
- Content triggers (e.g.,
DROP TABLE, api_key, private_key → L3)
- Configurable via
.failsafe/config/policies/risk_grading.json
Policy Evaluation: Each risk grade has deterministic requirements:
- L1: Heuristic check, 10% sampling, auto-approve
- L2: Full Sentinel pass, no auto-approve
- L3: Formal verification + human approval required
Ledger Recording: Every governance decision is recorded to an append-only SOA ledger with:
- Agent identity and trust score
- Artifact path and risk grade
- Timestamp and decision rationale
Trust Dynamics: Agent trust scores evolve based on outcomes:
- Approved L3 actions → trust increase
- Rejected or failed actions → trust decrease
- Trust scores influence future routing decisions
Why Deterministic Matters
When an LLM is asked to enforce safety rules, it can:
- Reinterpret rules based on context
- Produce inconsistent decisions across similar inputs
- Be influenced by prompt engineering attacks
QorLogic avoids these risks by executing deterministic TypeScript code at the governance boundary. The policy engine uses simple string matching and path analysis,no LLM inference required for governance decisions.
Example: A file containing api_key will always trigger L3 classification. No prompt can persuade the code to ignore this trigger.
Safety Alert
FailSafe policy denied editor save (write not prevented): AXIOM 1 VIOLATION: No active Intent exists.
Remediation: Create an Intent before modifying files. Run: failsafe.createIntent
[Create Intent] [View Active Intent] [Show Logs]
Features
1. Governance Modes
FailSafe now supports three governance modes to match your workflow needs:
| Mode |
Behavior |
Best For |
| Enforce |
Default. Full control, intent-checked saves (denials reported, not prevented), L3 approvals. |
Governed development, compliance |
| Assist |
Smart defaults, auto-intent creation, gentle prompts. |
Lighter-touch workflows |
| Observe |
No blocking, just visibility and logging. Zero friction. |
Exploration, learning |
Switch modes via:
- Command:
FailSafe: Set Governance Mode
- Settings:
failsafe.governance.mode
2. Save-Time Governance Gate
FailSafe evaluates save operations against the active Intent and reports a policy denial when no active Intent exists or when a file is out of scope. VS Code does not let an extension cancel an editor save, so the denial is observed, not prevented: the file may still be written.
3. Sentinel Monitoring and Audits
- File watcher queues audits for code changes
- Manual audits via command
- Modes:
heuristic, llm-assisted, hybrid (LLM uses the configured endpoint)
4. SOA Ledger and L3 Queue
- Append-only ledger database for audit entries
- L3 approvals surfaced in the UI
5. UI Screens
- FailSafe Monitor (compact view)
- FailSafe Command Center (extended popout/editor view)
- Skills view now includes
Recommended, All Relevant, All Installed, and Other Available to keep full skill visibility.
FailSafe Monitor UI

6. Command Center UX
- Compact
FailSafe Monitor webpanel (UI-02) provides phase status, prioritized feature counters, Sentinel state, and workspace health at-a-glance.
Open FailSafe Command Center opens the extended popout console for deeper workflow views (Overview, Operations, Audit, Risks, Skills, Laws, Mindmap, Config).
- Branding is consistent across shell surfaces, including FailSafe icon usage in header and favicon contexts.
UI Positioning Model
- Monitor and Command Center roles are defined in the Solution summary above; this model maps those roles to FailSafe architecture.
- Narrative alignment:
Genesis -> Build
QorLogic -> Govern
Sentinel -> Watch
Command Center -> Build + Govern
Monitor -> Watch
7. Skill Governance and Provenance
- Installed skills are discovered from FailSafe workspace roots (
FailSafe/VSCode/skills, .agent/skills, .github/skills) with project-first precedence.
- Phase-aware relevance ranking returns
recommended, allRelevant, and otherAvailable groupings.
- Skill metadata includes provenance fields (creator, source repo/path, source type/priority, admission state, trust tier, version pin).
SOURCE.yml metadata is ingested to preserve attribution and authorship for bundled and imported skills.
8. Checkpoint Reliability Backbone
- Checkpoint events are stored in a local SQLite ledger (
failsafe_checkpoints) with typed events and parent-chain integrity checks.
- Hub APIs expose checkpoint summaries and recent checkpoint history for UI transparency.
9. Feedback Capture
- Generate, view, and export feedback snapshots
10. QorLogic Propagation
Supported via internal sync flows when enabled by workspace governance configuration.
11. Break-Glass Protocol (v4.1.0)
Emergency governance overrides for time-sensitive situations. Activate via FailSafe: Activate Break-Glass Override with a justification (min 10 chars) and duration (15–240 minutes). Auto-reverts on expiry. Full audit trail recorded in the SOA ledger.
12. Verdict Replay (v4.1.0)
Re-execute past governance decisions for audit verification via FailSafe: Replay Verdict (Audit). Compares current policy hash and artifact hash against the original decision to detect drift.
13. Commit Governance (v4.3.0)
FailSafe: Install Commit Hook writes a thin hook client and per-session token into .git/.
- The hook calls
GET /api/v1/governance/commit-check and only enforces the server's allow decision.
- If the local API is unreachable or no token is present, the hook fails open by design. This is an operator guardrail, not a hard security boundary.
14. AI Provenance Tracking (v4.3.0)
- Save events can emit
PROVENANCE_RECORDED ledger entries with artifact path, detected agent type, confidence, and active intent.
- Provenance is observational. It does not mutate source files or inject comments.
- History is queryable through the governance API and visible in local ledger data.
15. Multi-Agent Governance Fabric (v4.2.0)
FailSafe detects and governs multiple AI coding assistants in your workspace:
- Runtime Detection: Identifies Claude CLI, Copilot, Codex CLI, and Agent Teams via terminal and config scanning.
- Per-Agent Config Injection: Writes governance rules into each agent's native format (
.github/copilot-instructions.md, .kilocode/rules/, codex.md, .claude/agents/).
- Governance Ceremony: Single command (
FailSafe: Set Up Agent Governance) to inject or remove governance across all detected agents.
- Coverage Dashboard: Console view showing which agents are detected, governed, and compliant.
- First-Run Onboarding: Guides new users through multi-agent governance setup on first activation.
16. Intent Schema v2 (v4.2.0)
Intents now carry schemaVersion, agentIdentity (which agent created the intent and via which workflow), and planId references. Legacy v1 intents are auto-migrated on read.
Commands
| Command |
Description |
| FailSafe: Open Console (Browser Popout) |
Main governance popout |
| FailSafe: Open Console (Browser) |
Browser launch alias |
| FailSafe: Open Console (Editor Tab) |
Compact monitor in editor |
| FailSafe: Token Economics Dashboard |
Open token economics and ROI dashboard |
| FailSafe: Audit Current File |
Manual file audit |
| FailSafe: Secure Workspace |
Apply workspace hardening baseline |
| FailSafe: Panic Stop |
Stop active monitoring and guard actions |
| FailSafe: Resume Monitoring |
Resume Sentinel monitoring |
| FailSafe: Set Governance Mode |
Switch between Observe/Assist/Enforce |
| FailSafe: Open Project Overview |
Project-level governance summary |
| FailSafe: Open Risk Register |
Open the risk tracking panel |
| FailSafe: Revert to Checkpoint (Time-Travel) |
Revert workspace to a governance checkpoint |
| FailSafe: Activate Break-Glass Override |
Emergency time-limited governance bypass |
| FailSafe: Revoke Break-Glass Override |
Manually revoke an active break-glass session |
| FailSafe: Replay Verdict (Audit) |
Re-execute a past governance decision |
| FailSafe: Undo Last Attempt |
Rollback to a specific checkpoint |
| FailSafe: Set Up Agent Governance |
Inject governance into detected AI agents |
| FailSafe: Install Commit Hook |
Add pre-commit governance hook to current repo |
| FailSafe: Remove Commit Hook |
Remove FailSafe pre-commit hook and token |
| FailSafe: Agent Health Status |
View composite agent health and risk level |
| FailSafe: Agent Execution Timeline |
Step-by-step agent action timeline |
| FailSafe: Shadow Genome Debugger |
Browse and debug failure patterns |
| FailSafe: Agent Run Replay |
Replay recorded agent execution traces |
Configuration
Open Settings and search for FailSafe:
| Setting |
Default |
Description |
failsafe.governance.mode |
enforce |
Governance mode: observe, assist, or enforce |
failsafe.genesis.livingGraph |
true |
Enable Living Graph visualization |
failsafe.genesis.cortexOmnibar |
true |
Enable Cortex Omnibar |
failsafe.genesis.theme |
starry-night |
Genesis UI theme |
failsafe.sentinel.enabled |
true |
Enable Sentinel monitoring |
failsafe.sentinel.mode |
heuristic |
Sentinel operating mode |
failsafe.sentinel.localModel |
phi3:mini |
Ollama model for LLM-assisted mode |
failsafe.sentinel.ollamaEndpoint |
http://localhost:11434 |
Ollama API endpoint, user settings only. A workspace may name only a local endpoint; another host is used after you allow it (FailSafe: Forget Sentinel LLM Approvals clears that) |
failsafe.sentinel.ragEnabled |
true |
Persist Sentinel observations to local RAG store |
failsafe.qorelogic.ledgerPath |
.failsafe/ledger/soa_ledger.db |
Ledger database path |
failsafe.qorelogic.strictMode |
false |
Block on all warnings |
failsafe.qorelogic.l3SLA |
120 |
L3 response SLA (seconds) |
failsafe.qorelogic.externalRuntime.enabled |
false |
Deprecated: has no effect. Remove this setting. |
failsafe.qorelogic.externalRuntime.baseUrl |
http://127.0.0.1:7777 |
Deprecated: has no effect. Remove this setting. |
failsafe.qorelogic.externalRuntime.apiKey |
`` |
Deprecated: not used by FailSafe. Remove any plain-text value. |
failsafe.qorelogic.externalRuntime.apiKeyEnvVar |
QORE_API_KEY |
Deprecated: has no effect. Remove this setting. |
failsafe.qorelogic.externalRuntime.timeoutMs |
4000 |
Deprecated: has no effect. Remove this setting. |
failsafe.bootstrap.autoInstallGit |
false |
Install Git system-wide if missing (asks before git init) |
failsafe.feedback.outputDir |
.failsafe/feedback |
Feedback output directory |
If .failsafe/config/sentinel.yaml exists, it overrides settings. The initializer seeds it with mode: hybrid unless you change it.
Commit-time governance is available through the local FailSafe API on http://127.0.0.1:7777 when the optional git hook is installed.
Workspace Files
FailSafe seeds a .failsafe/ directory in your workspace for configuration, ledger, and feedback output. The primary workspace config is .failsafe/config/sentinel.yaml. Optional policy overrides can be placed at:
.failsafe/config/policies/risk_grading.json
.failsafe/config/policies/citation_policy.json
Privacy
- Heuristic mode runs locally
- LLM-assisted and hybrid modes call the configured endpoint
Requirements
- VS Code 1.101.0 or later (the native SQLite module needs the Node 22 runtime VS Code ships from 1.101)
- Node.js 18+ (for development)
curl (required only if you install the commit hook)
- Ollama (optional, for LLM-assisted mode)
We'd love your review! If FailSafe is useful to you, please leave a review on the VS Code Marketplace or Open VSX. Your feedback helps other developers discover FailSafe and directly shapes its roadmap. Bug reports and feature requests welcome on GitHub Issues.
Contributing
Contributions are welcome via GitHub issues and pull requests.
Terms and Conditions (Beta)
FailSafe is a beta product. It is provided "as is" without warranties of any kind, and may contain bugs, incomplete features, or breaking changes.
By using this software, you acknowledge that it is experimental and agree to use it at your own risk. MythologIQ is not liable for any loss of data, downtime, or other damages arising from use of this beta release.
License
Apache License 2.0 - See LICENSE.
Links
Publishing
Releases are automated via GitHub Actions. Tag pushes trigger the CI/CD pipeline which builds, tests, and publishes to both VS Code Marketplace and Open VSX.
UI Snapshot

Checkpoint Integrity and Local Memory
FailSafe tracks more than Git state. It records governance checkpoints as signed metadata records, then stores Sentinel observations in a local retrieval store so operators can recover the what, why, and how of runtime decisions.
Process Reality
- Git readiness is checked at bootstrap (
ensureGitRepositoryReady). FailSafe asks before running git init in a folder without a repository, and installs Git only when failsafe.bootstrap.autoInstallGit is turned on.
- Governance events are checkpointed into
failsafe_checkpoints with run/phase/status context and deterministic hashes.
- Each checkpoint carries
git_hash, payload_hash, entry_hash, and prev_hash so chain integrity can be recomputed.
- Hub and API surfaces expose both summary and recent checkpoint records for operational visibility.
- Sentinel writes local memory records to
.failsafe/rag/sentinel-rag.db (or JSONL fallback), including payload_json, metadata_json, and retrieval text.
Technical Advantages
- Tamper evidence via hash-chained checkpoint records.
- Git-linked governance state for repository-correlated audit trails.
- Local-first memory retention for security and low-latency recall.
- Deterministic fallback paths when SQLite is unavailable.
Claim-to-Source Map
| Claim |
Status |
Source |
v4.3.0 ships commit hook install/remove commands. |
implemented |
FailSafe/extension/src/extension/main.ts, FailSafe/extension/package.json |
v4.3.0 ships commit-check and provenance API routes. |
implemented |
FailSafe/extension/src/api/routes/governanceRoutes.ts |
v4.3.0 exports governance context in release CI. |
implemented |
.github/workflows/release.yml, tools/export-governance-context.sh |
| Bundled operator docs ship inside the VSIX. |
implemented |
FailSafe/extension/.vscodeignore, FailSafe/extension/docs/COMPONENT_HELP.md, FailSafe/extension/docs/PROCESS_GUIDE.md |
Checkpoints persist in failsafe_checkpoints with typed governance fields. |
implemented |
FailSafe/extension/src/roadmap/RoadmapServer.ts:1533-1556 |
Checkpoint records include hash-chain material (payload_hash, entry_hash, prev_hash). |
implemented |
FailSafe/extension/src/roadmap/RoadmapServer.ts:1689-1695 |
| Each checkpoint captures current Git head/hash context. |
implemented |
FailSafe/extension/src/roadmap/RoadmapServer.ts:1647 |
| Checkpoint history and chain validity are exposed over API. |
implemented |
FailSafe/extension/src/roadmap/RoadmapServer.ts:331 |
Hub snapshot includes checkpointSummary and recentCheckpoints. |
implemented |
FailSafe/extension/src/roadmap/RoadmapServer.ts:742-743 |
| Sentinel local RAG persists observation payload + metadata + retrieval text. |
implemented |
FailSafe/extension/src/sentinel/SentinelRagStore.ts:60-81 |
| Sentinel RAG can fall back to JSONL when SQLite is unavailable. |
implemented |
FailSafe/extension/src/sentinel/SentinelRagStore.ts:85-91 |
RAG writes are controlled by failsafe.sentinel.ragEnabled (default true). |
implemented |
FailSafe/extension/src/sentinel/SentinelDaemon.ts:339-341 |
| Checkpoint and Sentinel RAG tables are independent (no foreign-key link). |
false |
Confirmed: failsafe_checkpoints (ledger DB) and sentinel_observations (RAG DB) are in separate databases with no shared keys. evidenceRefs is always []. |
| |