adaptogen-land

adaptogen-land

You are in the adaptogen-land state of the adaptogen DAG, always entered from adaptogen-critique with at least one landable finding in hand. This skill is self-contained.

What must hold here, always

• Never treat "found" as "fixed." A finding noted only in the adaptogen-critique report with no corresponding file edit here is observed, not improved. Every landable finding gets an actual edit before this state transitions onward.
• Edit the file the finding is actually in -- the target's router SKILL.md, a specific sibling state-skill's SKILL.md, a script or reference file beside either, or a project file elsewhere in the target's host repo. A finding about one state's wording gets edited in that state's own skill file, not bundled into a change to the router or a different state.
• Never remove a skill's capability from one run's evidence. Confirm the 2+ independent runs bar (checked in adaptogen-critique) before removing or replacing anything; a first occurrence adds a case, it does not delete one.
• A step proven deterministic becomes a script, not more prose. A step qualifies only when it has no judgment call in it ("decide" / "assess" / "pick") AND a prior run (checked via the target's own git log) already recorded it producing identical output from identical input. Write that step as a script in the relevant skill's own scripts/ directory and have that skill's text call the script instead of re-describing the steps in prose -- scripts/ is the agentskills.io-defined home for this, not a bespoke convention.
• Read every edit back to confirm it landed before moving to the next finding or transitioning onward. An edit that was written but not confirmed on disk does not count as landed.

This state also restructures target skills into DAGs -- concretely

A target skill that is a single flat SKILL.md is restructured into a DAG of sibling skills whenever this pass's evidence calls for it -- not reflexively on every pass (see the router's Gotchas: a target already working correctly as flat prose isn't split just for its own sake), but whenever a landable finding is specifically that the target has genuine states, branches, loops, or regressions that a single flat file is failing to hold onto correctly (a step being followed out-of-order, an invariant silently dropped on a non-linear path, a step's context assumed from a different section that a real run showed wasn't actually available at that point). When that finding lands, do this, not a lighter-weight alternative (a "states" section inside one file, states as references/ entries an agent optionally consults):

1. Identify the genuine states in the target's current procedure -- distinct concerns with their own invariants and their own outbound transitions (progression, loop, or regression) to other states. Do not invent states finer-grained than the target's real branches; a linear procedure with no branch stays one skill.
2. Rewrite the target's SKILL.md as a router: state map (an ASCII diagram naming every state-skill and the condition on every legal transition between them), an Entry section dispatching Skill(skill="<target>-<first-state>"), and the Hard Rules that bind across every state of the target.
3. Write each state as its own independent, agentskills.io-compliant skill directory, sibling to the router: <target-dir>-<state>/ containing its own SKILL.md with valid frontmatter (name matching the directory exactly), a description written for real standalone discovery (not just "state N of the DAG" -- another agent should be able to tell what this skill does from its description alone), and a body that: states what must hold in that state (repeated inline, not cross-referenced back to the router, since skills share no context); states what to do; states explicit transitions naming which sibling skill(s) to Skill()-dispatch next, the firing condition for each, and exactly what state to carry forward into that dispatch.
4. This is the same shape as adaptogen's own DAG (this very skill family) -- when in doubt about the right granularity or wording for a target's state-skill, look at how adaptogen-orient / adaptogen-run / adaptogen-critique / this skill are themselves written, since they are a worked example of the convention being applied.
5. Verify agentskills.io compliance on every new/changed skill directory before moving on: name frontmatter matches its directory exactly, lowercase-hyphen-only, no leading/trailing/consecutive hyphens; non-empty description; SKILL.md under ~500 lines (move detail to that skill's own references/, never back into the router or a sibling state's file).

This is the concrete instruction referenced by the router's Hard Rule 9 and the design principle stated there: accomplish what we're trying to describe instead of describing what we're trying to accomplish -- restructuring a target this way is not decoration on top of the fix, it is often the fix, when the evidence shows a flat file's prose failed to force the actual sequence a real run needed.

Do this (every pass, DAG-restructuring or not)

1. For each landable finding, in any order: open the specific file it lives in, make the edit, read the file back, confirm the edit matches what the finding required.
2. If a finding is out-of-scope (belongs to a different project/repo than the target being improved this pass), do not edit it here -- record it plainly in the pass's report as a finding for a separate, future pass on that other target. Landing an edit outside the current target's own repo in the same commit as the target's fix mixes two unrelated changes into one commit.

Transition

• All landable findings from this pass have a confirmed edit -> dispatch Skill(skill="adaptogen-commit-confirm") now. State explicitly: the list of edits made (file + what changed + which finding it addresses), the target's resolved absolute path and repo root, and (if this pass restructured the target into a DAG) the full list of new sibling skill directories created.
• No other transition exists from this state.