Skip to content

[pipeline] growth.yml saturates max-turns on 9/9 runs, produces 0 actions — cap/scope mismatch #176

@verkyyi

Description

@verkyyi

Source

14-day usage_log.md analysis (2026-04-18 CLI session).

Pattern

Every single growth.yml run in the window hit its max-turns cap:

  • runs: 9
  • max-turns cap: 25
  • at-cap hits: 9/9 (100%)
  • avg turns: 28.8 (exceeds cap — indicates pre-existing counter oddity or cap was recently raised)
  • result counter: "no_action" on every run (project_state.md 5th consecutive, confirmed across wider window)

Cost: $9.15 over 9 runs = $1.02/run. ~$5-10/wk while producing no progress.

Why this is a bug (not "plateau")

A saturating workflow that produces zero output is burning reasoning tokens on something. Two possible causes:

Hypothesis A — cap is too low: growth is discovering actionable work but running out of turns before it can act (e.g. opening a release candidate, drafting a discussion post, responding to stars). Raising cap would unblock it.

Hypothesis B — prompt scope is too broad: growth is scanning stars/forks/PRs/discussions/partner repos/etc. and burning turns on surveillance rather than action. The "5 consecutive no_action" pattern with full saturation is consistent with this — it's doing research instead of growth.

Hypothesis A is unlikely given current state: the 7 open needs-human growth items (#22, #48, #149, etc.) are blocked on external action, not on growth.yml capacity. So growth has nothing to DO even if it had more turns.

Proposed change (Hypothesis B)

Trim the growth.yml prompt to stop scanning things it can't act on:

  1. Skip repeat scans of needs-human items that have been held >48h without human action — log once, not every run
  2. Skip portfolio-level scans (agentfolio, etc.) unless evidence suggests cross-pollination opportunity
  3. Lower max-turns cap to 15 as a forcing function: if growth can't reach a decision in 15 turns, it's ruminating

Expected result: growth completes in 10-15 turns with the same "no_action" output when there's genuinely nothing actionable, at ~40% of current cost.

Acceptance criteria

  • growth.yml max-turns reduced to 15
  • Prompt scope narrowed (specifics TBD by coder based on current growth.yml prompt)
  • After 3 post-change runs: avg turns < 12, at-cap rate = 0%, no_action still the honest output when nothing is actionable
  • Cost per run target: < $0.40

Risk

Low. Worst case: growth misses a discoverable action one week that a human would have caught → one missed release candidate or discussion post. The current state is missing those anyway (5 consecutive no_action).

Related


Surfaced from 14-day retrospective CLI session 2026-04-18.

Metadata

Metadata

Assignees

No one assigned

    Labels

    likely-agent-fixableWatcher determined this pipeline issue can be fixed by the coder agentpipeline-fixPipeline failure to fix

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions