Skip to main content
← Back to list
01Issue
BugOpenSwamp CLIPublicEnterprise
AssigneesNone

Relationships

#4990 Swampify ranks swamp's own command sequences as candidates, and reads cost tokens

Opened by tagur · 11/22/2026

Summary Swampify's candidate set is dominated by swamp's own CLI calls. These are normal swamp usage, not chores from outside swamp. Reading them wastes reader tokens, and they crowd out real chores.

Evidence (30-day window, Claude Code, 35 candidates)

  • 13 of the top 20 candidates are call sequences made only of swamp commands (swamp behavior, swamp data, swamp workflow, swamp help, swamp extension, swamp vault).
  • Rank 11 (swamp behavior -> swamp extension: 15 occurrences, 9 sessions) and rank 19 (swamp help -> swamp extension: 9 occurrences, 6 sessions) are the same chore in two orders: find, pull and wire up an extension. Both show median 17384 weighted tokens and 2 tool calls per occurrence. They were not merged at the candidate stage, so each got its own reader (3 sessions each).
  • Readers judged most swamp-only sequences not-a-chore or low-value (the run-then-query loops, help lookups, vault lookups).

Impact

  • Reader tokens spent on candidates that are mostly discarded: 20 candidates read for 2 options.
  • Only 2 options survived from 20 reads, one of them a swamp-only sequence.

Suggested

  1. Merge candidates that differ only in step order or in a swamp help step, before reading.
  2. Down-rank or skip sequences made only of swamp commands. The user already works through swamp there. Cover them as one swamp-usage group, or ask the user first.
  3. Steer the swampify skill: tell it to prefer chores outside swamp (git, curl, deno, other CLIs) in the first reading batch.
  4. Optionally cap reading by token budget, not by candidate count.
02Bog Flow
◉OPEN○TRIAGED○IN PROGRESS○SHIPPED

Open

11/22/2026, 1:16:55 AM

No activity in this phase yet.

03Sludge Pulse
tagur moved this from a private collective to public11/22/2026, 1:17:36 AM

Sign in to post a ripple.