Repository navigation
Pilot a versioned coordinator and subagent workflow #8989
Description
Goal
Pilot a small, versioned agent-orchestration workflow for LibreSign after the reusable skills and isolated development environment are available.
Runtime dependency
Use a runtime that provides genuinely separate subagent contexts. The current reference implementation is the OpenAI Agents API multi-agent capability. Do not emulate subagents with repeated sequential prompts or duplicated analysis in one context.
This dependency is independent from Behat and from LibreSign's test-suite modernization.
Approach
Start with one coordinator that can delegate independent work to temporary subagents with separate context.
Do not create permanent specialized agents for every engineering role.
Candidate parallel tasks include:
- independent investigation of an issue;
- review of a completed diff with clean context;
- targeted test/compatibility research that does not depend on another subtask's intermediate state.
Sequential work whose next step depends on the previous output should normally remain in one execution flow.
Execution environments
Use hosted execution for ordinary repository work where it is sufficient.
Use self-hosted/persistent execution only when the task needs capabilities such as:
- the canonical Docker/Nextcloud environment;
- multiple isolated worktrees;
- long-lived Docker image/cache reuse;
- browser/integration tests requiring a real local service topology.
Self-hosting must not become a mandatory transfer hop for all work.
Versioning
Keep coordinator instructions/configuration in version control with explicit inputs, allowed tools, handoff rules and evaluation cases.
Do not add WORKFLOW.md or SPEC.md merely for convention. Introduce a workflow/spec file only if the chosen orchestrator actually consumes it.
Evaluation
Compare the pilot with a single-agent flow on representative tasks, recording at least:
- completion correctness;
- human interventions required;
- wall-clock latency where measurable;
- duplicated/repeated work;
- tool failures/handoffs;
- token/model cost where available.
Done when
- The coordinator workflow is versioned and reproducible.
- Subagents are used only for genuinely independent work.
- Hosted versus self-hosted execution has a concrete decision rule.
- Representative trials show whether orchestration reduces manual coordination without adding a new bottleneck.
Activity
vitormattos commentedon Oct 5, 2026
Started the first orchestration pilot increment in LibreSign/agent-skills#14.
The implementation deliberately reuses
execute-issueas the coordinator instead of adding a new orchestration framework. Single-flow execution remains the default; temporary subagents are only used for genuinely independent investigation. The skill also defines hosted vs self-hosted decision rules, isolated-write constraints, and evaluation metrics for comparing coordinated and single-flow runs.The agent-skills validation workflow is green on the PR.
This is an incremental implementation of #8989 rather than completion of the whole issue; representative single-agent vs coordinator trials are still needed before closing it.
vitormattos commentedon Oct 5, 2026
Current remaining gap
The versioned coordinator policy is implemented in
LibreSign/agent-skillsand the execution-environment decision rule is in place.What remains for this issue is narrower than the original scope: run representative real single-flow vs coordinator + temporary-subagent trials and record correctness, intervention, latency/duplication, handoff failures and cost when exposed.
This current ChatGPT tool environment does not expose a primitive for spawning genuinely separate subagent contexts, so I am not treating sequential tool calls or repeated analysis as a valid substitute. Fabricated or simulated orchestration metrics would not satisfy the issue.
No additional orchestration framework, permanent role agents, WORKFLOW.md or SPEC.md is needed before that capability exists.
vitormattos commentedon Oct 5, 2026
The remaining runtime dependency is now concrete: OpenAI's Agents API provides real independent subagent contexts and explicit multi-agent orchestration.
Implementation is in LibreSign/agent-skills#18. It extends the existing
execute-issueworkflow instead of adding another skill, and adds a reproducible read-only runner that executes the same task packet in single-flow and multi-agent modes, recording wall-clock time, real subagent creation events, output, and usage when exposed.This is unrelated to Behat. #8986 and this issue have both been updated to make that boundary explicit.
After #18 is merged, the only remaining step for #8989 is to run the controlled pilot with Agents API access and score the resulting correctness/intervention fields.