fix(using-superpowers): re-check skills mid-workflow, not only at turn start

[arimu1]

This PR MUST target the dev branch, not main. main is the
released branch; active work lands on dev first. PRs opened against
main will be asked to retarget dev before review.

Who is submitting this PR? (required)

Field	Value
Your model + version	Grok 4.5
Harness + version	Grok Build / Cursor
All plugins installed	Superpowers skill pack in use for contribution workflow; no third-party skill plugins installed for authoring this change. Skill text edits hand-written against issue #2051 (not auto-generated skill rewrite).
Human partner who reviewed this diff	arimu1

What problem are you trying to solve?

Real session reported in #2051 (Claude Code + Superpowers 6.2.0, Claude Sonnet 5).

using-superpowers states that a relevant skill should be checked before any action, with even a 1% chance of relevance enough to invoke it. In the reported session:

1. brainstorming engaged correctly for a rate-limiting design.
2. Mid-conversation, design settled with an explicit rejected alternative and a second non-obvious decision — content that matched a second skill's trigger description.
3. Implementation proceeded; the second skill never activated.

When asked why, the agent explained (verbatim from the issue):

"Once brainstorming's design questions wrapped up, I was tracking the brainstorming skill's own terminal state... not scanning for other skills that might apply to what had just been decided. The using-superpowers rule says to invoke a skill if there's even a 1% chance it's relevant — I applied that at the start of the turn but not again after the design conversation produced exactly the kind of rejected-alternative/rationale content [the other skill] exists to capture."

Root instruction gap: the bootstrap forbids skipping skill checks before action, but does not say the check is continuous. Agents rationalize "I already checked this turn" / "I'm mid-workflow so skill check is done."

A third-party skill cannot fix this from its own body: trigger decisions run against short frontmatter descriptions; body text only loads after the skill is already chosen. Only using-superpowers can restate continuous re-check.

Issue author noted prompt-based enforcement may not hard-guarantee compliance; this PR closes the documented wording gap and the exact rationalizations observed.

What does this PR change?

Skill-content only in skills/using-superpowers/SKILL.md:

1. The Rule: state that skill check is continuous (not only at turn start); mid-workflow does not suspend checking other skills; when a second skill's trigger is met, invoke it (respecting Skill Priority).
2. Red Flags: two rows closing the observed rationalizations ("already checked this turn"; "mid-workflow — skill check is done").

No harness/hook/code changes. No third-party skill content.

Is this change appropriate for the core library?

Yes. Continuous skill discovery is core bootstrap behavior for every Superpowers user on every project. It is not domain-specific, not a third-party integration, and not a project-local workflow. Third-party skills cannot fix this from their own files (as #2051 demonstrated).

What alternatives did you consider?

1. Only add Red Flags rows — Leaves The Rule readable as once-per-turn-start. Rejected: the positive rule must state continuity; Red Flags alone are the rationalization counter.
2. Only strengthen The Rule without Red Flags — Rejected: using-superpowers' "check any skill before every action" doesn't hold once a workflow step is underway #2051's agent produced exact mid-workflow rationalizations; project pattern is to table those (see prior "I know what that means" row for explicit-request skips).
3. Change skill description frontmatter — Rejected: description is discovery/when-to-load; this skill is session bootstrap. Body is where continuous behavior is enforced. Description already says "before ANY response."
4. Hook/runtime enforcement of re-scan — Out of scope for skill-content fix; no portable hard guarantee across harnesses; issue asked about the stated policy gap.
5. Larger rewrite of bootstrap / flowcharts — Rejected: bootstrap was deliberately compressed (token cost on every session); keep the minimal wording that blocks the observed failure mode.

Does this PR contain multiple unrelated changes?

No. Single file, single failure mode: skill check treated as turn-start only while mid-workflow.

Existing PRs

• I have reviewed all open AND closed PRs for duplicates or prior art
• Related PRs: none found for issue using-superpowers' "check any skill before every action" doesn't hold once a workflow step is underway #2051 or for "mid-workflow skill re-check / continuous skill check while one skill is active"

Searched open and closed PRs for: 2051, using-superpowers, mid-workflow, check continuously, before every action. Nearby using-superpowers PRs (#2010 dangling anchor, #1894 Claude Code platform adaptation, harness PRs) do not address continuous re-check. Related issues cited by #2051 (#1007 composition, #1812 memory) are different problems.

Environment tested

Harness (e.g. Claude Code, Cursor)	Harness version	Model	Model version/ID
Grok Build / Cursor	Cursor IDE (current)	Grok 4.5	Grok 4.5

Skill-content-only change; no runtime harness integration. Grounding session is the #2051 transcript (Claude Code / Claude Sonnet 5).

New harness support (required if this PR adds a new harness)

N/A — this PR does not add a new harness.

Clean-session transcript for "Let's make a react todo list"

N/A — not a new harness PR

Evaluation

• Initial prompt / source of truth: Issue using-superpowers' "check any skill before every action" doesn't hold once a workflow step is underway #2051 session (Claude Code + Superpowers; brainstorming active; second skill trigger met mid-design via rejected alternative + non-obvious decision; second skill never self-triggered).
• Before (bootstrap wording + observed agent behavior):
  • Rule: invoke before any response/action — read as once at turn start
  • No continuous / mid-workflow re-check language
  • Agent tracked active skill's terminal state only; did not re-scan other skill descriptions when new trigger content appeared
• After (this PR wording):
  • Explicit: check continuously, not only at turn start
  • Mid-workflow does not suspend other-skill checks
  • Second skill with met trigger: invoke (Skill Priority still applies)
  • Red Flags name the two rationalizations from using-superpowers' "check any skill before every action" doesn't hold once a workflow step is underway #2051
• Adversarial / pressure checks (reasoning + diff review against issue transcript; skill text change only):
  1. Exact using-superpowers' "check any skill before every action" doesn't hold once a workflow step is underway #2051 case: After brainstorming produces rejected-alternative content matching another skill's description → continuous re-check + "mid-workflow" red flag forbid treating skill check as done.
  2. True single-skill turn: No new triggers mid-turn → no extra invocation required; rule is "when trigger is met," not "invoke everything every message."
  3. Skill Priority still holds: Process skills first; continuous check does not invert priority or force parallel exclusive workflows.
  4. SUBAGENT-STOP unchanged: Subagents still ignore this skill; no change to that gate.
  5. Token budget: +~90 words on a deliberately compressed bootstrap; no flowchart restore; no section restructuring.
  6. Carefully-tuned content: Existing Red Flags rows left as-is; two additive rows only. No "human partner" rewording. EXTREMELY-IMPORTANT / 1% block unchanged.
• Multi-session live harness evals: Not run in this contribution environment (skill markdown only). Prompt-based enforcement remains soft (as using-superpowers' "check any skill before every action" doesn't hold once a workflow step is underway #2051 author noted); this closes the documented gap. Before/after multi-session evals remain welcome as follow-up if maintainers want quantified behavior rates.

Rigor

• If this is a skills change: I used judgment aligned with superpowers skill philosophy (precise labels, forbid the observed rationalization, keep carefully-tuned sections intact). Full writing-skills multi-subagent campaign was not run in this harness; adversarial cases above map the using-superpowers' "check any skill before every action" doesn't hold once a workflow step is underway #2051 failure and adjacent loopholes.
• This change was tested adversarially, not just on the happy path (exact using-superpowers' "check any skill before every action" doesn't hold once a workflow step is underway #2051 mid-brainstorming miss, single-skill no-op, Skill Priority, SUBAGENT-STOP)
• I did not modify carefully-tuned content (existing Red Flags rows, rationalizations, "human partner" language) without extensive evals — existing rows left unchanged; only additive continuous-check rule + two new rows matching using-superpowers' "check any skill before every action" doesn't hold once a workflow step is underway #2051 rationalizations

Human review

• A human has reviewed the COMPLETE proposed diff before submission

Fixes #2051