Skip to content

Replace dedicated sub-agents with skills - #82

Merged
linted merged 1 commit into
mainfrom
worktree-skills-instead-of-subagents
Nov 18, 2026
Merged

linted merged 1 commit into
mainfrom
worktree-skills-instead-of-subagents

Conversation

@linted

@linted linted commented Nov 18, 2026

Copy link
Copy Markdown
Owner

The coordinator used to delegate to 8 dedicated sub-agents (program-recon, function-analyst, research, general-purpose, prototype-fixer, type-fixer, vuln-hunter, switch-fixer), each with its own prompt, tool allowlist, model and write tier. This replaces that with a coordinator + single generic worker, where the analysis/repair playbooks live as on-demand deepagents skills.

  • skills/: 7 SKILL.md playbooks (analysis/: program-recon, investigate, vuln-hunt; repair/: function-analysis, prototype-recovery, type-recovery, switch-recovery), read on demand via progressive disclosure.
  • Coordinator holds the full tool set and applies small changes itself; it delegates verbose/parallel work to one general-purpose worker, naming the skill in the brief.
  • Write tiers collapse to per-mode: normal mode writes everything; plan/ask drop every MUTATION_TOOLS entry (and the repair skills) from both coordinator and worker, read-only by construction. Removed the annotations tier, the maybe_ prefix enforcement and the pending-change queue.
  • subagents.py -> agent_config.py ([main]/[worker] models only); subagents.toml -> agents.toml. Skills mounted via a CompositeBackend /skills/ route over the working backend, with a deny-write permission.
  • Tests updated; new skills/config guards in test_shipped_config.py validate every SKILL.md through deepagents' own loader and catch tool-name drift.

The coordinator used to delegate to 8 dedicated sub-agents (program-recon,
function-analyst, research, general-purpose, prototype-fixer, type-fixer,
vuln-hunter, switch-fixer), each with its own prompt, tool allowlist, model and
write tier. This replaces that with a coordinator + single generic worker, where
the analysis/repair playbooks live as on-demand deepagents skills.

- skills/: 7 SKILL.md playbooks (analysis/: program-recon, investigate,
  vuln-hunt; repair/: function-analysis, prototype-recovery, type-recovery,
  switch-recovery), read on demand via progressive disclosure.
- Coordinator holds the full tool set and applies small changes itself;
  it delegates verbose/parallel work to one general-purpose worker, naming the
  skill in the brief.
- Write tiers collapse to per-mode: normal mode writes everything; plan/ask drop
  every MUTATION_TOOLS entry (and the repair skills) from both coordinator and
  worker, read-only by construction. Removed the annotations tier, the maybe_
  prefix enforcement and the pending-change queue.
- subagents.py -> agent_config.py ([main]/[worker] models only);
  subagents.toml -> agents.toml. Skills mounted via a CompositeBackend /skills/
  route over the working backend, with a deny-write permission.
- Tests updated; new skills/config guards in test_shipped_config.py validate
  every SKILL.md through deepagents' own loader and catch tool-name drift.
@linted
linted merged commit c360a63 into main Nov 18, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant