Changelog

Versions are git tags. VERSION in the working tree names the release; install.py status prints both that and the actual checkout, so a mid-line or modified clone is visible rather than silently claiming to be a release.

See INSTALL.md for upgrading, pinning and rolling back.

v0.15.1 — 2026-10-30

• Makes the Git source native-install-ready by shipping generated per-skill shared references; the package builder and regression tests reject missing/stale copies of canonical resources.
• Uses one stable Git-backed vince-gate marketplace and vince-gate@vince-gate identity. Documents register-once installation, same-marketplace upgrades, tag pinning and rollback.
• Adds real isolated Codex install/upgrade/rollback proof using two local Git fixture versions; verifies host tool calls, installed resources, exactly one marketplace and preserved project data.
• Updates README, INSTALL, user guide and CLI/plugin docs together; distinguishes native Python 3.11+ requirements from legacy/manual installation and corrects the stale six-skill introduction.
• Preserves 0.15.0 release assets and normal user configuration; documents migration from its versioned local marketplace source. No new review rules or additional product tools.

v0.15.0 — 2026-10-30

• Packages a desktop-first native plugin with portable and Codex compatibility manifests, bundled skills and shared references, scripts, templates, icons and local MCP configuration.
• Adds four read-only local tools: plugin info, project context, task status and structural ledger validation. Explicit workspace roots, bounded reads and no arbitrary shell executor.
• Review checks the stated requirements and relevant tests, writes one verdict, and stops. Removes default audit quotas and automatic remediation/re-review from caller instructions.
• Adds reproducible ZIP/checksum/release-note builds, clean-install smoke tests and release CI.
• Preserves the legacy installer and Claude plugin. Lifecycle hooks are not included in the native directory ZIP; standalone opt-in hooks remain available from the source repository.
• Fixes an expired routing-test fixture by supplying its existing deterministic test date.
• Requires Python 3.11+ and Git. No cloud service; public directory publication is subject to OpenAI publisher verification and desktop-MCP submission eligibility, not this GitHub release.
• Release baseline is main at 0161e79 plus the scoped review policy ported from 93b0cc2. Historical first-review/readiness experiments and their unresolved review status are not shipped.

v0.14.0 — 2026-09-12

• Adds vince-brief: a compact, task-persistent response contract for action-first Vince updates.
• Every Vince workflow now applies the shared brief contract only when forming its user-facing result.
• The contract preserves detail on request and never substitutes concise presentation for safety, authorization, evidence, or a known cause.

v0.13.0 — 2026-09-09

• Profiles and lessons now use compact, budgeted rule formats to reduce context cost.
• doctor --fix sanitizes every discovered Vince workspace/repo, preserves custom content, creates content-addressed backups, and is idempotent.
• Vince Learn writes operational one-line lessons instead of narrative review prose.

v0.12.0 — 2026-08-31

Vince adopts progressive disclosure across its primary workflows and adds native Agent Skills bindings for current Gemini CLI and GitHub Copilot layouts.

Changed

• Every Vince skill ends user-facing work with concise Result, optional Problem, and Next lines; detailed proof remains in task artifacts.
• Review Pass 0 freezes a validated exhaustive coverage manifest, and discovery continues across all material claims and A0–A7 attacks even after FAIL is certain.
• vince-implement and vince-review keep their proof floor in compact primary instructions; phase-specific contract, TDD, wire-proof, attack, verdict, and handoff details now load from focused references only when needed.
• Shared voice and token-discipline references are conditional instead of mandatory on every skill activation, while their essential safety and efficiency rules remain inline.
• Discovery descriptions are distinct and capped by regression tests, and both primary workflows stay below a conservative approximate 5,000-token activation budget.

Harnesses

• Gemini CLI now renders native skills to .gemini/skills/ and ~/.gemini/skills/ instead of legacy TOML custom commands.
• GitHub Copilot gains native .github/skills/ and ~/.copilot/skills/ bindings without sharing Codex's install directories.
• Both new/changed bindings remain explicitly unverified until live discovery and invocation probes pass in their runtimes.

Verified

• Token, reference, description, binding, documentation, and disposable-install regression tests cover the progressive-disclosure contract.
• Existing routing, intake, Codex compatibility, external security boundary, and Stop-hook tests remain in the full suite.

v0.11.2 — 2026-08-31

Vince adds bounded operational helpers and hardens shipped skill instructions while preserving its existing gate architecture.

Added

• Health reporting. The operations CLI reports binding, task, route, and next-action health.
• Route refresh. Exact model inventories can refresh routing profiles without guessed IDs.
• Release checks. Version, changelog, tag, and disposable-install consistency can be checked before publishing.
• Codex discovery probe and task archiving. Maintainers can verify live skill discovery and archive only tasks whose current verdict is PASS and whose workspace is clean.

Security

• Shipped skill examples use project-local executables instead of unpinned package runners.
• Regression coverage prevents external audit tooling from becoming an embedded Vince command, installer gate, baseline, or shipped-skill dependency.
• Shipped skill Markdown rejects package-runner tokens across platform suffixes, paths, wrappers, prose, and Markdown fence styles.

Verified

• The full suite passes with 34 tests and no skips.
• Independent adversarial review passed after mutation-testing CLI boundaries, shipped-skill coverage, Windows executable variants, and alternate Markdown representations.

v0.11.1 — 2026-08-23

Vince now routes implementation work to the least expensive capable model class and the narrowest useful agent role, while keeping the proof floor intact.

Added

• Model routing gate. vince-route classifies work as economy, balanced, frontier, or reviewer, then resolves exact harness-specific model identifiers from the project profile.
• Agent role routing. Profiles can map exact explorer, worker, and reviewer roles per harness, with no provider-name fallback.
• Switch recommendations. Vince can recommend an exact model switch with the token/quality tradeoff, while making clear that the harness or user performs the actual switch.

Verified

• Codex live routing passed a nine-case matrix covering trivial, standard, explorer, complex, security, reviewer, switch, ask, and proof-floor behavior.
• Stale, unavailable, unchecked, missing, or unverified mappings return ASK instead of silently substituting another model.
• Claude, Gemini, Cursor, Windsurf, and generic bindings render the routing skill and profile fields, but remain marked as render-verified rather than live-verified.

v0.11.0 — 2026-08-17

Codex support is now based on live runtime behavior rather than configuration shape alone.

Fixed

• Native reviewer role. The supplied TOML now uses Codex's required developer_instructions; the prior description-only role was rejected by the runtime.
• External verdict persistence. Reviewer sessions whose Vince task directory is outside the repository now document the required parent --add-dir <resolved task-root> grant.
• Stop-hook task lookup. The hook finds both repository-local and external Vince task stores, including repositories without an origin remote.

Verified

• Project-scoped Vince skills were installed, discovered and invoked by Codex CLI 0.148.0-alpha.9.
• A persistent Codex parent delegated to the named reviewer role, which ran the installed vince-review skill and persisted a genuine verdict in an external task directory.
• Project .codex/hooks.json did not load, and user-scope Stop blocking was not exercised. The binding and documentation state that limitation instead of claiming hard enforcement.

Upgrade notes

Copy the updated reviewer template into .codex/agents/vince-review.toml. Use a persistent Codex parent session, and when install.py where --repo <repo> resolves outside the repository, launch Codex with --add-dir <resolved task-root>.

v0.10.0 — 2026-08-14

Two bindings were wrong, found by checking them against current documentation instead of trusting what they were written from.

Fixed

• Codex. Was installing to .codex/vince/ with no frontmatter and an AGENTS.md pointer block. Codex has had native skill loading since December 2025: .agents/skills/ for a project, ~/.agents/skills/ or ~/.codex/skills/ for a user, most specific first, with the same name + description frontmatter Claude Code uses. The canonical skills now install unchanged and auto-activate, and the pointer block is no longer needed.
• Cursor. Reference docs were being written as plain .md into .cursor/rules/, which Cursor ignores - no frontmatter means it is not a rule, so they sat in the rules directory looking registered while doing nothing. Bindings gained a reference_dir, so references now install to .cursor/vince/ and links inside each rule are rewritten to ../vince/…. The rules directory contains only real rules. Same fix applied to Windsurf.

Added

• A per-harness capability table in docs/harnesses.md. The compatibility question is not whether files install, it is whether the harness can give the reviewer a fresh context. Claude Code and Codex can, natively. Cursor, Windsurf and Gemini CLI cannot, so the review is a second chat you open and paste the handoff into - the same rigour with one manual step, and arguably cleaner isolation, but easier to skip. Stated plainly rather than implied.
• templates/codex-reviewer-agent.toml - a Codex subagent definition for the review. Codex subagents are TOML in .codex/agents/ and can pin the model, which is the one place the reviewer's model can actually be chosen rather than merely recommended.

Upgrade notes

Anyone with a previous Codex install should uninstall the codex binding before reinstalling; the old .codex/vince/ location and its AGENTS.md block are no longer written, and uninstall only removes what its manifest recorded.

v0.9.0 — 2026-08-13

Makes "you can clear context mid-task" a checked claim instead of a hopeful one, and stops the docs implying Vince can do things it cannot.

Added

• scripts/resume.py - rebuilds a task's state from its ledger alone, reading no conversation. Prints what a fresh session needs (phase, next action, open criteria with their proof commands, live resources still to tear down) and returns SAFE TO CLEAR / NOT SAFE TO CLEAR with the specific gaps. A ledger missing a Resume block, a proof command on an open criterion, or a baseline is not self-sufficient, and it says so rather than letting you find out after clearing.
• Checkpoint protocol. At every phase boundary: bring the ledger current, update its new Resume block, run resume.py --check. Never suggest clearing without that check passing - a reset on an incomplete ledger destroys work, which is worse than the context it saves.
• Pressure signals - proxies an agent can actually observe, since it cannot see its own token count: ~15 files read or ~5 suite runs since the last checkpoint, a large diff or build log captured, the same file read more than twice, entering remediation. Documented as approximations.
• checkpoints profile setting: off (default), suggest, insist. At suggest, Vince offers a /compact at checkpoints that pass the safety check. It cannot run compaction itself - that is the user's keystroke - and the skills say so explicitly rather than implying otherwise.
• Model honesty. A skill cannot select its own model. Vince now states which model it is running as in the verdict and handoff, passes reviewer_model/mechanical_model through so whoever spawns the subagent can honour them, and recommends the split - strongest model for the judgement passes, cheaper for search and mechanical sweeps. It does not claim to have chosen.

Upgrade notes

checkpoints defaults to off, so nothing changes until you set it. Existing ledgers gain a Resume block the next time a task touches them; resume.py --check will flag its absence.

v0.8.0 — 2026-08-13

Token discipline. Vince was directly implicated in its own users' cost profile: it mandates a subagent per task, its handoff recommended general-purpose by name, and vince-implement alone was ~10k tokens of context on every task. Rigour is not negotiable; what it costs is.

Added

• scripts/check.py - the deterministic half of a review as a script instead of model turns. One command reports stray artifacts, AI/bot attribution trailers, over-long commit subjects, newly added skipped tests, debug statements in non-test code, possible hardcoded secrets, whole-file rewrites (the line-ending-flip shape), a branch behind its base, an uncommitted tree, and whether any test files were touched. Replaces roughly ten commands and their raw output. Explicitly input, not verdict - it proves nothing about behaviour, isolation or blast radius.
• skills/_shared/token-discipline.md - read narrowly, bound long commands, spend the model where judgement is needed, and treat the ledger as memory so context can be reset. Ends with what never gets cut to save money.
• Profile settings reviewer_agent_type and mechanical_model, and tokens on the metrics line so vince-learn can report cost per task against rounds-to-PASS.

Changed

• vince-implement 10,098 -> ~7,400 tokens. Remediation, self-healing, commit hygiene and workspace hygiene moved to reference/remediation.md and reference/hygiene.md, which load at the step that needs them; the inline ledger example is replaced by a pointer to the template. What stayed inline is everything that changes behaviour every task.
• The reviewer handoff no longer names general-purpose as the default. Use the narrowest agent type with Write/Edit; general-purpose is the fallback, because a broad type carries a bigger prompt for capabilities a review never uses.
• vince-review A1 now starts with the script and falls back to hand commands.
• voice.md trimmed - it is copied into all eight skills, so its examples were paying eight times.

Upgrade notes

No required config. reviewer_agent_type and mechanical_model are optional. check.py needs nothing installed - it is in the toolkit clone.

v0.7.0 — 2026-08-12

Added

• A voice. Vince is now dry and a little sarcastic, and it explains its own jargon instead of assuming you speak it - the precise term stays, with the plain-English translation beside it, including a glossary of the toolkit's own vocabulary (surviving mutant, wire proof, blast radius, isolation key, first-touch promotion...).

  Three rules stop the personality costing anything. A joke never carries information - strip every one out and the message is still complete. Comedy is in the delivery, never the verdict; severity and PASS/FAIL are always flat. And it is never aimed at the user: fair targets are the situation, the tooling and itself. It goes completely straight for security findings, anything destructive, and any time it was wrong - a joke there reads as dodging, because it usually is.

  Artifacts stay professional. Ledgers, verdicts, commit messages and completion docs are written plainly whatever the voice setting, because someone reads them later without the conversation around them.

• voice profile setting: playful (default), plain (same honesty, no jokes), terse (facts only). Asking in conversation is also a standing instruction.

• skills/_shared/ - reference content copied into every skill at install time. One source of truth in the repo, present next to each skill at runtime; a file eight skills each kept a copy of would drift within a release. _shared has no SKILL.md so it is never listed as a skill, and flat layouts render it as <skill>-voice.md with links rewritten.

Upgrade notes

No config fields are required. Add voice: to a profile only if you want something other than the default. Existing installs pick the voice file up on the next install.

v0.6.1 — 2026-08-12

Fixes the discovery gap that 0.6.0 shipped with, found by the incident that prompted the skill: two python -m http.server processes, started a week earlier, holding a dist/ directory open. 0.6.0's primary sweep filtered Win32_Process on CommandLine and ExecutablePath containing the directory - and neither field contains it. The association is in the working directory and open handles. The sweep returned nothing and the directory looked unheld.

Changed

• vince-cleanup now sweeps on three axes instead of a path filter: listening ports (a preview server exists to serve, so it is bound to something), age + image name (dev tooling up for days is nobody's active work), and open handles. Per-candidate resolution then answers what the command line cannot - handle.exe -p, lsof -p ... cwd, readlink /proc/<pid>/cwd.
• Added a table of shapes that hide from a path grep: python -m http.server, npx serve, php -S, ruby -run -e httpd, bare node dev servers, dotnet watch, kubectl port-forward.
• Post-stop verification: process gone, port released, directory writable. "Killed it" is not "the directory is free" - the incident had two servers on the same port, so killing one changed nothing observable.
• "Nothing is holding it" now sends you back through the three-axis sweep before concluding the directory is unheld, rather than straight to Explorer windows and antivirus.
• New attack-playbook entry (A5 sweep): a stale server on the wire-proof's port answers instead of the build under review, with week-old output. A process older than the branch cannot be serving the branch.

Upgrade notes

No config changes.

v0.6.0 — 2026-08-12

Added

• vince-cleanup. Recovers a workspace after a session ended without tearing down: leaked git worktrees, processes still holding directories open, background jobs nobody stopped, stray build and scratch output. vince-implement already told a session to sweep up after itself and vince-doctor reported orphaned worktrees, but neither handles processes and neither helps once the session that made the mess is gone.

  It is the only skill that kills processes and deletes directories, so it is built to refuse: inventory first, attribute every item as yours / unknown / someone else's and act only on the first, never kill by process name, never --force past a dirty or unpushed worktree, never rm -rf a worktree at all. It also owns the diagnosis nothing else had - which process holds a directory open when a remove fails (handle.exe / Win32_Process, lsof / fuser), worked in escalation order instead of reaching for force.

• A Session resources block in the ledger. One row per worktree and long-running process, written as it is created rather than at the end, where a crashed session never reaches it. This is the attribution that lets cleanup act confidently instead of asking about everything.

• The Stop hook catches leaked worktrees: a ledger reading PASS whose recorded worktree is still on disk means teardown was skipped. Only PASS ledgers are checked, and a path that is gone or still a template placeholder is never a leak.

Upgrade notes

No new profile fields. Existing ledgers keep working; the Session resources block is additive and only affects tasks started after the upgrade. If you use the Stop hook, it now blocks on leaked worktrees as well as unproven rows - VINCE_STOP_DISABLE=1 still turns it off entirely.

v0.5.0 — 2026-08-11

Vince no longer writes into the repos it works on. 0.3.0 made per-repo profiles effectively mandatory - the reviewer FAILs a baseline taken from an unverified inherited command - and then put them in <repo>/.vince/, which meant an untracked directory in every work repo. Verification is per-repo by nature and that does not change; where the record is stored is independent of what it describes, and that is what was wrong.

Added

• A config store outside the repos. Per-repo profile, lessons, metrics and task ledgers live at <store>/repos/<key>/, defaulting to ~/.vince/ and overridable with $VINCE_STORE (point it at <workspace>/.vince to keep an estate's config with its hub profile).
• Remote-derived repo keys. github.com__acme__billing-api, from the origin remote - stable across re-cloning and moving a checkout, readable enough to find and hand-edit, and shared by task worktrees, so a -wt worktree resolves to its parent repo's config instead of a blank one. Repos with no remote fall back to local__<name>__<short path hash>.
• install.py where [--repo DIR] [--json] - prints the key, mode and every resolved path. Path resolution is code, not model judgement, so every session agrees and the store cannot fragment. The skills now call it instead of deriving paths.

Changed

• Per-repo config is in the store by default. A repo that carries its own .vince/profile.md still wins, which is the opt-in for repos you own and want the profile committed for a team.
• Nothing to gitignore in store mode, because nothing lands in the repo.

Upgrade notes

Existing in-repo .vince/ directories keep working untouched - they are the opt-in path now. To move a repo's config out, move <repo>/.vince/ to the path install.py where --repo <repo> prints once the in-repo profile.md is gone, then drop the gitignore entry. No config fields were added or renamed.

v0.4.0 — 2026-08-11

Added

• vince-update. Moves an install between releases, and owns the half install.py cannot: reading the changelog between the installed and target versions, stopping on uncommitted toolkit work / in-place edits / a task in flight rather than forcing past them, reinstalling at the scope and bindings the manifest recorded, and then migrating .vince config so a new release's fields exist instead of being silently absent. New fields arrive unknown or blocked with what would resolve them - never invented. Rollback is the same flow with an older tag, and deliberately does not strip config, since older skills read newer profiles fine.

Changed

• The release checklist now requires actionable upgrade notes naming every config field or section a release adds, because vince-update reads them to migrate existing projects.

Upgrade notes

No new config fields. vince-update itself installs like any other skill; after upgrading, use it instead of git pull + reinstall so config migration happens too.

Workspace profiles, for estates with many repos under one hub. Driven by a real limitation: a hub cannot run a hundred suites, so a hub-level profile that claims verified commands is lying.

Added

• Two-level profiles. <workspace>/.vince/profile.md carries the repo map, stack definitions, branch model, tracker, isolation key, environments and estate-wide gates; <repo>/.vince/profile.md carries verified commands and the observed baseline.
• The invariant, stated everywhere it is read: a hub profile cannot verify a command. Every per-stack command in a hub profile is (inferred, unverified) by construction, and confidence does not survive inheritance.
• First-touch promotion. The first task in a repo runs what it inherited, writes that repo's profile with the commands that actually worked and the baseline actually observed, and records any hub default that was wrong. It costs nothing extra - the task was running them anyway.
• Merge semantics. Scalars override; dod_extras and known_traps are additive and the hub's cannot be removed by a repo; lessons are read at both levels and routed by scope.
• Section status vocabulary: verified, inferred, unverified, unknown - <tried>, blocked - <what is needed>. Wire-proof rigs and mutation tooling are usually legitimately blocked at onboarding; a blank section reads as nobody thought about it, blocked reads as a gap with an unblock.
• vince-setup gains workspace mode; vince-doctor validates both levels and reports hub profiles claiming impossible verification, plus repos worked in that never got a profile; vince-learn routes lessons by scope.
• Per-repo baselines on multi-repo tasks. The ledger carries one row per repo in dependency order. A suite never run in repo B cannot tell you whether you broke repo B - and a baseline taken with an unverified inherited command is now a reviewer FAIL condition.
• templates/workspace-profile.template.md.

Upgrade notes

Nothing breaks. A single-repo profile keeps working unchanged and needs no Inherits from header. Adopt the hub model by running vince-setup at the workspace root; existing repo profiles are already in the right shape.

v0.2.1 — 2026-08-11

Version selection. v0.1.0 and v0.2.0 shipped without git tags, so there was no way to install or roll back to a specific release - only "whatever main happens to be". Both are now tagged retroactively at the commits that were those releases.

Added

• Annotated tags v0.1.0 and v0.2.0.
• install.py prints the toolkit's git ref beside VERSION, so 0.2.1 (v0.2.0-4-gabc-dirty) shows a mid-line or modified clone instead of letting it claim to be a clean release.
• This changelog, and a Versions section covering what-am-I-running, upgrade, pin, roll back, agent-instruction pinning, and holding the Claude Code plugin at a tag via extraKnownMarketplaces with a ref.

v0.2.0 — 2026-08-11

Makes the review measure rather than judge. Three changes aimed at the same weakness: a review resting on the model's opinion of work the model just did.

Added

• Blind first pass (vince-review Pass 0). The reviewer derives findings from the diff and the original contract before opening the ledger, completion doc, commit messages or any prior verdict — all of which it now treats as claims, not evidence. The verdict records how many findings came blind versus only after. Skipping Pass 0 is a hard FAIL condition.
• Tool-backed mutation testing. The profile gains a mutation section naming the stack's tool and its diff-scoped invocation (StrykerJS/.NET, mutmut, PIT, go-mutesting, cargo-mutants). TAMPER and the reviewer's A2 both run it; a mutant surviving on a changed line is a missing assertion to kill or waive, and a hard FAIL if neither. Hand mutation remains the fallback.
• reviewer_model. The profile can name a different model to review with. Fresh context breaks the correlation introduced while generating, not the correlation in a model's parameters.
• Claude Code plugin packaging. /plugin marketplace add elroykanye/vince-gate, then /plugin install vince-gate@vince-gate. Skills arrive namespaced: /vince-gate:vince-implement.
• Opt-in Stop hook (hooks/vince_gate_stop.py) that blocks a session from ending while the active ledger has unproven rows or no PASS verdict. Fails open in every ambiguous case.
• New DoD gate: new code survives mutation.
• install.py reports the toolkit's git ref alongside VERSION.

Changed

• --binding auto now includes generic whenever the project has an AGENTS.md, rather than using it only as a last-resort fallback. Most coding agents read AGENTS.md, so it is the widest-reach binding available.
• README states the positioning directly: spec-driven frameworks enforce process; this enforces evidence.
• The docs are more explicit about limits — independent review still finds a minority of defects and barely improves on contextual errors, which is why the wire proof and mutation gate carry as much weight as the reviewer.

Upgrade notes

Nothing breaks. Existing .vince/profile.md files keep working; the new mutation and reviewer_model fields are optional and both skills degrade to the previous behaviour without them. To adopt them, re-run vince-setup or add the sections by hand. Reinstall with install.py install against the same scope — the manifest records what that was.

v0.1.0 — 2026-08-11

First release.

• Six skills: vince-setup, vince-implement, vince-review, vince-document, vince-doctor, vince-learn.
• Harness bindings for Claude Code, Cursor, Windsurf, Codex, Gemini CLI and any AGENTS.md runtime, rendered from one canonical markdown source. Claude and generic verified.
• install.py with install / status / doctor / uninstall / list / bindings, a checksummed manifest, drift detection, and refusal to clobber in-place edits.
• Templates for the profile, verification ledger, review verdict, lessons and completion doc.
• Paste-to-your-agent install guide.

Rewritten to be stack-agnostic from a set of skills originally built for one platform.

---

Releasing

1. Land the work on main.
2. Bump VERSION in the same commit as the last change of the line — in 0.2.0 the bump landed last, so three commits carried 0.2.0 features while claiming 0.1.0. install.py now prints the git ref for exactly this reason, but the bump still belongs with the work.
3. Update this file. Upgrade notes are not optional and must be actionable: name every config field or section the release adds, because vince-update reads them to migrate existing projects. "Nothing breaks" is a complete note only when nothing was added.
4. Bump version in .claude-plugin/plugin.json — plugin users only receive updates when that string changes.
5. Tag annotated and push: git tag -a vX.Y.Z -m "…" && git push origin vX.Y.Z.