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

Relationships

#4706 Swampify: the skill recognises both swampify briefs, and compare measures raw against swamp in a compared state

Opened by skunk-ape · 11/13/2026· Shipped 11/13/2026

Part of Swampify onboarding. Builds on #4642 (the swampify skill and the offer, pick and built methods) and #4650 (comparison v2 in core's contract). The kick-off brief comes from #4705 (swamp start).

Repo: swamp-extensions (@swamp/agent-sessions).

Goal

Two things, in this order:

  1. The swampify skill recognises both swampify briefs that core writes, so swamp start and the Swampify page reach it cleanly.
  2. Swampify measures what it saved. A compare method writes a comparison for each built option, and the skill shows it in a compared state.

Part 1: brief entry

Core writes every start brief under the one-line header <!-- swamp start brief v1 -->, with no kind: field. So the skill's current entry, a header with kind: assess, never fires.

The skill should recognise the two swampify briefs by their content:

  • Kick-off (from swamp start). It holds a window and a scope and tells the agent to pull the extension, create the swampify behavior and follow the skill. The skill takes the window and scope, confirms them back in one line, and starts at disclosed, skipping its questions.
  • Build (from the Swampify page, swampifyBuildBrief). It names an offer behavior and an option id. The skill follows the brief's own steps: record the pick, build through its hand-off, run once, record built. Then it continues at compared and graduated.

Also:

  • Fix the line "any other brief belongs to getting-started" so it doesn't contradict a build brief. Only a brief with no Swampify content belongs to getting-started.
  • Agree the exact recognisable wording with #4705, and pin it with a test on both sides.

Part 2: compare

A compare method takes an offer option that has reached built, and writes or rewrites swampify-comparison-<option id> (comparison v2):

  • Build cost. The tokens and active time of the session that built the option. That is the session whose calls ran this offer's pick and built, found from session records.
  • History. Past occurrences of all the option's candidates, summed across them, with per-occurrence cost as two ranges:
    • chore: the occurrence's own calls;
    • wholeTurns: the whole turns that contained it.
  • Swamp run. Wall-clock time, model tokens (zero for a behavior run) and outcome of the first run of the target, from swamp's own run records.
  • Forward. From when it was built onward:
    • raw: occurrences still done by hand, meaning new occurrences of the option's candidate ids in a later assess;
    • swamp: occurrences done through the target, meaning session calls that ran the built behavior or workflow, from session records.
  • Break-even. The occurrences that recover the build cost, low against whole turns and high against the chore's own calls, or null when it never pays back.
  • Money. Leave out estimatedUsd unless a price table is supplied. When it is shown, it is labelled an estimate.

compare reads records and candidate sets only. It calls no model and re-reads no transcripts.

Part 3: the skill's compared state

  • After built: run compare and show raw against swamp as ranges, never as one number, with the build cost and break-even always shown. Then go to graduated.
  • On a later visit, re-run import, assess and compare, so the forward figures fill in.
  • Repeatability as a fact: the swamp side gives the same output each run; the raw runs don't.

Out of scope

  • Replay of a past occurrence.
  • The Swampify page, which already shows a comparison when one exists.
  • Telemetry.

Done when

  • The skill's brief recognition has tests for kick-off, build and non-Swampify briefs, matched to core's kick-off wording.
  • compare has unit tests for:
    • finding the build session;
    • history ranges summed across merged candidates;
    • forward attribution, both raw and through swamp;
    • break-even, including never paying back;
    • a target that has no run yet.
  • A property test shows that chore never exceeds wholeTurns, and that every comparison passes core's reader.
  • The skill passes tessl review at 90% or above, and its trigger and routing evals still pass.
  • A manual run, from a swamp start kick-off through a build to compared, shows a comparison in the terminal and on the Swampify page. The implementer checks this locally, and no transcript content goes in the PR.
02Bog Flow
✓OPEN✓TRIAGED✓IN PROGRESS✓SHIPPED+ 1 MOREASSIGNED+ 5 MOREREVIEW+ 19 MOREPR_MERGED+ 2 MORESESSION_SUMMARIZED

Shipped

11/13/2026, 1:39:19 PM

Click a lifecycle step above to view its details.

03Sludge Pulse
skunk-ape assigned skunk-ape11/13/2026, 3:06:00 AM
Editable. Press Enter to edit.

system commented 11/13/2026, 3:00:23 AM

Classified automatically when this issue was filed.

  • Source: Extensions

If you feel this classification is incorrect, add a ripple to tell us so.

Sign in to post a ripple.