Skip to content

Self-Improvement

Self-improvement (SI) is an optional issue-seeding loop. An enabled track periodically creates an issue that asks Colony to find improvement candidates in a repository. It does not bypass analysis, development, or review. By default a person selects each candidate; a track can select automatically only when its selection mode and the repository’s autonomy grant allow it.

Each stored track has a name, label, enabled state, cadence, proposal slate, weight, and active prompt version. The prompt version provides:

  • the title and body of the seed issue
  • instructions the analyzer uses to identify candidate improvements

The proposal slate specifies the candidate sizes and counts to offer—for example, two small candidates and one medium candidate. Weight sets the track’s share of its pool: it decides how many open slots the track gets and favors under-share tracks when Colony picks the next seed. It does not change priority in the shared task queue.

Seeding runs only when self-improvement is enabled, a track store is available, and the track itself is enabled. Before creating anything, Colony checks the track’s open linked SI issues against its slots, and skips a track whose latest seed is still waiting for candidate selection. A track’s slots come from its share of its pool within self_improvement.max_open_issues, with a minimum of one.

For cooldown cadence:

  • the first seed can be created immediately when there is no completed issue
  • later seeds wait until cooldown_minutes has elapsed since the track’s latest seed or latest completion, whichever is later

An open issue uses one of the track’s slots regardless of elapsed time. Canceling a linked SI issue records completion for this purpose.

For cron cadence:

  • the first seed can be created immediately
  • later seeds are eligible when the most recent UTC occurrence of the cron expression is later than the track’s latest seed or completion
  • the track still needs a free slot

After a track’s first seed, an invalid or missing cron expression does not seed.

When eligible, Colony:

  1. Loads the track’s active prompt version, with configured or built-in fallback text.
  2. Creates a real issue using that stored title and body.
  3. Persists the pipeline issue and track link as the authoritative records.
  4. Applies the track label to the code-host issue on a best-effort basis.
  5. Runs analysis using the track instructions and proposal slate.

The seeder does not dynamically fill the seed body with recent repository signals. Repository inspection happens during analysis.

The analyzer posts a candidate list for selection and waits for clarification. A human selects a candidate or requests another slate, unless the track selects automatically. After selection:

  • small or medium work can proceed to development
  • large work returns through analysis for decomposition
  • epic work enters planning and decomposition

From there, the selected work uses the same development, automated review, merge policy, and recovery paths as other pipeline issues.

Tracks can be managed directly. The optional Strategist can also propose changes to track status, cadence, weight, proposal slate, and prompts, but its snapshot requires human approval. Strategy proposals do not themselves seed an issue.

See Strategy and Configuration Schema for the two configuration layers.