Skip to main content
← Back to list
01Issue
FeatureShippedSwamp CLIPublic
Assigneesskunk-ape

Relationships

#4807 Swampify: guide the last mile from built to everyday use (an adopted state)

Opened by skunk-ape · 11/16/2026· Shipped 11/16/2026

Part of Swampify. Follows #4714 (one build per session, hand back to the page) and #4755 (built only for accepted output), from a first build-brief run.

Repo: swamp (the bundled swampify skill: SKILL.md and references/handoff.md, plus an eval).

Problem

Swampify ends at "built and measured", but at that point the automation isn't used anywhere yet. In a run that built "Stage and commit changes" as a @swamp/git behavior (stage-and-commit), everything that made it usable day to day happened only because the user asked afterwards:

  1. graduated stops right after the build. After a build brief, the skill prints one directive line and stops, with no next steps. Nothing prompts wiring the build into the user's flow.
  2. The build hard-codes its first-run target. handoff.md pushes toward one safe first run, so the agent wrote the scratch repo path and a demo author identity into globalArguments. The chore happens in many repos (842 times in 262 sessions), so the target needed to be an inputs field (repoPath: ${{ inputs.repoPath }}).
  3. Agents work outside the workspace. The user's agents run in their dev repos, not in the swampify workspace. Calling the behavior from there needs --repo-dir / SWAMP_REPO_DIR, which the skill never mentions.
  4. Agents are never told to use it. Without a rule in CLAUDE.md / AGENTS.md (global or per repo), agents keep doing the chore by hand.
  5. The adoption signal goes unused. compare already records forward.raw against forward.swamp, but the skill presents them only as numbers that fill in later.

Proposed solution

Add an adopted state between compared and graduated:

  • Where the chore happens. Ask one question, seeded from the candidate's session paths (the repos or projects it showed up in). If there's more than one target, turn the fixed targets into behavior inputs and test one run from outside the workspace with --repo-dir.
  • Remove first-run setup, such as scratch paths and demo identities, from the definition.
  • Tell the agents. Offer to add a short rule with the exact command to the user's global or per-repo CLAUDE.md (Claude Code) and AGENTS.md (Codex). Read the file first, write only on a yes, and leave each harness's attribution trailer as a placeholder. Mention a PreToolUse hook as an option for enforcing it, but don't push it.
  • After a kick-off or build brief: one short question ("Wire it into your agents? Where does this chore happen?"), then the same directive line as now. This keeps the page's one-build-per-session model.
  • Later visit: if forward.raw.occurrences is still above forward.swamp.occurrences, revisit the wiring before suggesting the next option.

handoff.md: when a chore spans targets (repos, accounts, clusters), say to make those targets inputs, using the scratch target only for the first run.

Eval: a build-brief run ends with the adoption question before the directive line, and a multi-target chore ends up as a behavior with an inputs field.

Alternatives

  • Leave it to the user to ask. That's what happened here, but most users won't think to.
  • Have the Swampify page show copy-paste wiring snippets under Built. That helps, but the agent is better placed to edit the instruction files and check the result.
02Bog Flow
✓OPEN✓TRIAGED✓IN PROGRESS✓SHIPPED+ 1 MOREASSIGNED+ 2 MOREREVIEW+ 7 MOREPR_MERGED+ 2 MORESESSION_SUMMARIZED

Shipped

11/16/2026, 9:16:56 PM

Click a lifecycle step above to view its details.

03Sludge Pulse
skunk-ape assigned skunk-ape11/16/2026, 7:45:34 PM

Sign in to post a ripple.