You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Workflow bundles: parameterized, shareable automations with one-click launch links #275
Upgrade automations from private schedules to reusable assets: export/import a workflow as a versioned bundle (fits #164's WORKFLOW.md), parameterize it (repo, branch, model, schedule), and launch via solenta:// deep link (#186). A community/team catalog of bundles is the stretch goal.
Research note (Xata Agent): include plain-English playbooks as a bundle type — ordered-step markdown procedures ('investigate failing CI: check logs, classify the failure, ...') that constrain the executing agent turn by turn. Users write their own; constrains hallucination better than a freeform prompt. Related: Anthropic's tool-testing agent idea — an automation that exercises flaky MCP tools and rewrites their descriptions, 40% faster completion for future agents.
Research note (marketplace round): adopt Claude Code's decentralized marketplace format for workflow bundles — a marketplace is any git repo with a manifest JSON. No server to run; versioning, discovery, and install come free from git. Skip paid marketplace mechanics entirely (GPT Store is the cautionary precedent); free templates as lead-gen.
Research note (AAIF Agent Plugins 1.0, Aug 6): the industry-default portable layout is plugin.json + skills/ + mcp.json + vendor-namespaced extras, installable in VS Code, Cursor, Codex, ChatGPT, Kiro. Adopt this layout for workflow-bundle export; note there is no provenance layer yet, so installs should be treated as code execution.
Deep-dive detail (Conductor API cookbook): ship literal copy-paste system prompts that turn any external agent into a fleet manager (plan, brief, spawn, supervise, report; multi-PR fanout; daily report). It is marketing, docs, and product in one. Note their constraint — 'the session shares no context with you' — which our shared memory explicitly counters; say so in the cookbook.
From the Goose doc-mine: recipe retry with shell success checks — after agent completion, declared checks run (external commands, not the model's say-so), and on failure the recipe runs cleanup, RESETS message history to initial state, and re-runs up to max_retries. Declarative verification-loop-in-config for workflow bundles. Also: /recipe generates a bundle from the current conversation — capture-what-worked as a one-command export.
Closing as not planned in the 2026-10-09 board cleanup (#1532). Speculative idea with no work and no tie to the current roadmap. Reopen if it becomes a priority.
Activity
currentbits commentedon Aug 17, 2026
Research note (Xata Agent): include plain-English playbooks as a bundle type — ordered-step markdown procedures ('investigate failing CI: check logs, classify the failure, ...') that constrain the executing agent turn by turn. Users write their own; constrains hallucination better than a freeform prompt. Related: Anthropic's tool-testing agent idea — an automation that exercises flaky MCP tools and rewrites their descriptions, 40% faster completion for future agents.
currentbits commentedon Aug 17, 2026
Research note (marketplace round): adopt Claude Code's decentralized marketplace format for workflow bundles — a marketplace is any git repo with a manifest JSON. No server to run; versioning, discovery, and install come free from git. Skip paid marketplace mechanics entirely (GPT Store is the cautionary precedent); free templates as lead-gen.
currentbits commentedon Aug 18, 2026
Research note (AAIF Agent Plugins 1.0, Aug 6): the industry-default portable layout is plugin.json + skills/ + mcp.json + vendor-namespaced extras, installable in VS Code, Cursor, Codex, ChatGPT, Kiro. Adopt this layout for workflow-bundle export; note there is no provenance layer yet, so installs should be treated as code execution.
currentbits commentedon Aug 18, 2026
Deep-dive detail (Conductor API cookbook): ship literal copy-paste system prompts that turn any external agent into a fleet manager (plan, brief, spawn, supervise, report; multi-PR fanout; daily report). It is marketing, docs, and product in one. Note their constraint — 'the session shares no context with you' — which our shared memory explicitly counters; say so in the cookbook.
currentbits commentedon Aug 18, 2026
From the Goose doc-mine: recipe retry with shell success checks — after agent completion, declared checks run (external commands, not the model's say-so), and on failure the recipe runs cleanup, RESETS message history to initial state, and re-runs up to max_retries. Declarative verification-loop-in-config for workflow bundles. Also: /recipe generates a bundle from the current conversation — capture-what-worked as a one-command export.
currentbits commentedon Oct 9, 2026
Closing as not planned in the 2026-10-09 board cleanup (#1532). Speculative idea with no work and no tie to the current roadmap. Reopen if it becomes a priority.