Release v0.14.0

What's Changed

Added

• Direct tool-run API for IDE clients (no LLM loop): POST /api/tools/run, GET /api/tools/run/{job_id}, DELETE /api/tools/run/{job_id} (cancel + SAQ abort), GET /api/tools/run-allowlist. Allowlisted tools (linter_run, test_runner, docker/git helpers, …) route via tool_routing.yaml to the native tools worker (SAQ) or run in-process. Config tools_run.allowed_tools plus AGENTFORGE_TOOLS_RUN_ALLOW / AGENTFORGE_TOOLS_RUN_DENY. See docs/api.md and docs/tools.md
• Playbooks: curated command-combinations + Jinja2 output templates for @discover. Registry playbooks.yaml + templates under markdown/playbooks/*.jinja, retrieved semantically from a dedicated Qdrant collection (agentforge_playbooks, threshold-gated). A match seeds the investigation's probe commands and replaces LLM scoping (single pass, skip_analysis), and the rendered template pre-structures the synthesis input. Per-playbook synthesis: diagnostic|informational picks the cleanup-plan or descriptive-report prompt. CLI: python -m agentforge.playbooks.cli reindex|list|match. Config under playbooks.*
• Recap: POST /api/sessions/{id}/recap returns a cumulative running summary of a session. Only messages after the previous recap are read (its sequence is the watermark) and the previous recap text seeds the prompt, so the summary covers the whole session at incremental cost. Persisted as a volatile recap message so it survives reloads without entering model context. Idle-safe: fewer than recap.min_new_messages (default 2) or no new messages returns the previous recap with no LLM call. Config under recap.*
• @rebase skill (markdown/skills/git-rebase.md): drives the local rebase-onto-default-branch flow, proposes per-file conflict resolutions and waits for approval, and enforces --force-with-lease (never bare --force, never the default branch)
• AgentForge Email knowledge collection: knowledge.mail_collection_name (kb_mail_entries), allowed via X-Knowledge-Collection / session source=mail. Search accepts projects: [...] (MatchAny) for multi-account All Inboxes filtering
• JetBrains / IntelliJ custom-agent prompt (markdown/custom-agents/intellij.md) for IDE-context agent runs

Changed

• Session compaction and recap share one conversation flattener (web/server/summarise.py) instead of each carrying its own copy
• Dockerfile.web ships playbooks.yaml alongside skills.yaml
• New dependency: jinja2 (playbook templates)
• linter_run: multi-tool quality groups (ruff/flake8, black/isort checks, …); tool_name=black|isort|flake8|mypy|… resolves via a built-in known-tool map when missing from config; improved --fix for black/isort
• Secret redaction: high-entropy catch-all defaults off (fewer CamelCase false positives); skip pure CamelCase / snake_case identifiers when HE is on; clearer SECRET_REDACTION_ENABLED force on/off; compose + native tools-worker launchd plist default redaction off for whole-file coding paths
• deploy-remote.sh health probes: retry loop, web checks /api/health (not /), correct compose container name prefix, non-fatal SSH/health failures so the Done banner still prints

Fixed

• Command policy store no longer migrates on the shell/SSH hot path. Opening the chat DB ran create_tables() (a full Alembic upgrade) on every command, which serialised each one behind a SQLite write-lock upgrade (journal_mode=DELETE) and deadlocked when the native worker and the web container share web_chat.db. A genuinely missing table is now repaired lazily by get_runtime_override; YAML policy stays authoritative either way
• @discover no longer double-persists discovery.plan (plan widget rendered twice on session reload)

Full changelog: https://github.com/bulletinmybeard/agent-forge/blob/master/CHANGELOG.md