Skip to content

Strategy

The optional Strategist assesses a repository’s charter goals against what Colony has recorded, and proposes changes to the strategy-owned self-improvement tracks. It does not write code or edit issues. Unless you switch approval to automatic, a person approves a proposal before it applies, and only then do the tracks change.

Each run, scheduled by Colony or requested on demand, collects:

  • the active charter: ordered goals and guidance
  • the live tracks and each one’s 30-day scorecard
  • signals such as open findings, pipeline outcomes, cost concentration and activity by area
  • repository memory and intelligence

A collector that fails is recorded as unavailable with its reason, and the run goes on without it. The run invokes the Strategist once, with repair attempts if its output fails validation, and stores one of three outcomes: a proposed snapshot, no change, or a failure with its reason. A run fails if the repository has no active charter.

A charter is an ordered list of goals, in priority order. Each goal has a one-sentence statement, an optional “Done when” and a stable id (g1, g2, …). Ids are never reused. The charter also holds guidance: free text the Strategist keeps in mind but never assesses.

For every goal, each run records an assessment: met, partly, unmet or unknown, with the evidence it rests on. Evidence cites its sources by prefix: scorecard:, signal:, memory:, intel: and file:. Every change a proposal makes cites what it read in the same way. The assessment is LLM output and can be wrong, which is why each claim names its source. Runs that find nothing to change still update the assessments, so the trend of a goal is visible without a proposal.

Self-improvement work runs in a limited number of open slots. Two pools divide them:

  • the strategy pool, for tracks the Strategist owns
  • the operator pool, for tracks you own

The repository’s capacity share (0 to 1, default 0.5) sets how many slots go to the strategy pool. With five open slots and a share of 0.5, three go to strategy tracks and two to operator tracks. Inside the strategy pool each track gets slots in proportion to its share, which a proposal sets as the track’s weight. Operator tracks are never changed by a strategy.

A strategy-owned track has a name, a share, a cadence (cooldown or cron), a proposal slate (how many candidates of each size analysis should offer), a prompt, and the goals it serves. A proposal can create tracks, change existing strategy tracks and retire them. Each directive names the goals it serves and cites its reasons. A track serving a goal that is already met is refused, as is a directive with no evidence. Your cooldown floor and selection mode apply to strategy tracks and are never overwritten by a proposal.

Size limits live in the repository’s autonomy grant, separate from the charter:

FieldCurrent behavior
max_auto_sizeHard limit on the largest proposed track slate size: small, medium, large or epic. Defaults to small.
max_auto_select_sizeThe largest candidate size that may be selected automatically. none keeps every selection with a person. Defaults to none.

A snapshot is created proposed. There is no resting approved state: approving applies the proposal at once.

StatusMeaning
proposedWaiting for review.
activeApplied. It is the strategy the repository runs now. Only one is active.
supersededReplaced by a newer active version, or set aside because the charter changed.
rejectedDeclined, with a reason.
invalidA plain approve was refused by validation. The violations are kept.

Colony checks every proposal against the same rules at approval time, so a narrowed grant refuses an older proposal. Approval is recorded with who decided: a person, or auto when approval is automatic. Revert applies the previous version’s tracks again as a new version. Work already started keeps going.

Colony’s own scheduler is the only one. It runs on an anchored cron (by default Mondays 03:00 UTC, staggered per repository), runs again 10 minutes after a charter change, and runs on request. You can switch scheduled runs off per repository; a request still works.

Strategy changes the tracks that self-improvement uses on later seeding cycles. Self-improvement then creates seed issues according to each enabled track’s cadence and requires a candidate-selection step before implementation.

See Self-Improvement for that execution path, Strategy for the review screen, and Configuration Schema for supported setup fields.