Changelog

1.3.0 - 2026-10-26

Learned from the first real run: developing an issue end to end on a production repo (plan, research wave, implementation, verification wave with a fresh build, adversarial pass, report). The phase structure earned its place; two friction classes did not survive it, and both are now closed.

• Workers apply file edits themselves and reply with the diff (worker.toml rule 7, run skill Phase 2): in the run, a worker proposed JSON that the orchestrator transcribed, the transcription lost the file's escape convention (3 vs 2 backslashes) and caused a fix round with two failed attempts.
• Mandate template in Phase 2 (repo/branch, ONE question, numbered specifics, constraints, reply format): the first run improvised it and it worked; written down, it is repeatable.
• Heavy-action workers (builds, measurements) must report side effects (worker.toml rule 8): in the run, a worker's fresh build regenerated tracked artifacts that no mandate covered, caught only at git diff --stat time.
• Phase 6 now requires a working-tree hygiene pass before the report: git status --short, every modified file classified (intended / side effect / pre-existing), and a mandatory "Side effects" section in the report.
• Collection loop leans on runtime notifications; agents.list polling is for overdue replies only (the run spent sleep+list rounds before the first reply had even arrived).
• Phase 2 asks for wave start times in findings.md and Phase 6 for total wall-clock in the stats: the cost warning becomes measurable.
• README: disambiguation section — this plugin is not Mistral's "Workflows" durable-orchestration platform in Studio (same name, different thing).

1.2.0 - 2026-10-26

Fixes for components that never loaded on Vibe 3.0.x, and a working install path for the enforcement hooks. All found via inspect_plugin: 1.1.0 (and 1.0.0) reported five load errors.

• Skills: both SKILL.md descriptions contained unquoted ": " and failed YAML frontmatter parsing, so workflows:run and workflows:stress-test never registered. Descriptions are now quoted and load cleanly.
• Agents: the three TOMLs used camelCase (schemaVersion, agentType, displayName, enabledTools) while the plugin schema requires snake_case (schema_version, agent_type, ...), so workflows:worker/verifier/adversary never existed. Converted.
• Hooks: Vibe 3.0.x parses plugin hooks but never executes them (the runtime consumes only user/project hooks.toml). install.sh now merges the three enforcement hooks into ~/.vibe/hooks.toml under workflows:-prefixed names, idempotently (skip with WORKFLOWS_NO_HOOKS=1). Verified end to end: a fresh session denies git push --dry-run via the shell guard.
• install.sh no longer copies the plugin's .git (read-only git object files broke re-installation).

1.1.0 - 2026-10-26

Native enforcement via plugin hooks (ai.mistral.vibe/hooks.toml).

• wf-spawn-budget (pre_tool on agents_spawn): denies workflows:* spawns once the session has used its budget — WORKFLOWS_SPAWN_CAP (default 40). Generic spawns are never blocked or counted.
• wf-spawn-audit (post_tool on agents_spawn): counts only spawns whose body actually ran and appends an audit line to ${VIBE_HOME:-~/.vibe}/workflows-hooks/spawn-audit.log, giving Phase 6 run stats for free.
• wf-shell-guard (pre_tool on shell tools): denies commands whose command position carries rm -rf-style flags, git push, git reset --hard, or git clean -f — defense in depth for the agent denylists, covering the orchestrator session and generic subagents. Mentions inside arguments (e.g. grep 'rm -rf' README.md) pass.

1.0.0 - 2026-10-25

First public release.

• workflows:run skill: full dynamic workflow with 7 phases (intake, plan with confirmation gate, wave-based fan-out, independent verification, adversarial pass, convergence loop with hard 3-round budget, coordinated report).
• workflows:stress-test skill: multi-angle plan stress-testing (feasibility, correctness, security, cost, alternatives) with per-section verdicts (KEEP / FIX / RETHINK).
• Subagents: worker (narrow mandates, FACT/HYPOTHESIS labeling), verifier (independent re-verification, CONFIRMED/REFUTED/UNCERTAIN), adversary (BREAKER/RISK/NIT attacks).
• Scratchpad checkpointing per run id (plan, findings, verification, adversarial, report) with resume-on-restart semantics.
• Straggler protocol: 3 sleep rounds, then a "reply now" nudge, then drop (failed workers are not fatal).
• Knowledge folder with the original-feature research and mapping.
• English docs, MIT license, installable as a user-global or project-scoped Vibe plugin.

0.1.0 - 2026-10-25

Initial private prototype: single run skill, three subagents, Italian docs.