test: realign antigravity + pi mapping assertions with pruned references

[gaurav0107]

Targets the dev branch.

Who is submitting this PR? (required)

Field	Value
Your model + version	Claude Opus 4.8 (1M context) — model ID claude-opus-4-8[1m]
Harness + version	Claude Code 2.1.143
All plugins installed	superhuman@superhuman, superpowers@claude-plugins-official
Human partner who reviewed this diff	Gaurav Dubey (@gaurav0107)

What problem are you trying to solve?

The plugin-infrastructure test suite fails on dev (and on main, 6.1.1). Two per-harness mapping tests assert content that was intentionally deleted:

$ git checkout dev
$ bash tests/antigravity/test-antigravity-tools.sh
FAIL: mapping does not document view_file as the file/skill-read tool   # exit 1

$ node tests/pi/test-pi-extension.mjs
✖ pi tools reference documents pi-specific mappings
  AssertionError: The input did not match the regular expression /write/   # exit 1 (1 of 6 cases)

Root cause: commit e7ddc25 ("Prune per-harness tool-mapping boilerplate", 2026-06-24) deliberately removed the skill-loading explainers and the generic action→tool tables from antigravity-tools.md and pi-tools.md, trimming each file "to the harness-specific notes that still carry weight (subagent dispatch, task tracking…)". That commit did not touch tests/, so the two content-assertion tests still grep for tokens the prune removed — view_file, IsSkillFile, run_command, grep_search (antigravity) and read/write/edit/bash (pi). With no CI in the repo, the breakage went unnoticed.

What does this PR change?

Updates the two orphaned tests to assert only the harness-specific mappings e7ddc25 kept (subagent dispatch and task tracking) instead of the boilerplate it removed. No reference, skill, or product content is touched — only the stale test assertions (and their now-inaccurate header comment / PASS string).

Is this change appropriate for the core library?

Yes. These are the repo's own plugin-infrastructure tests under tests/. No new dependency, no skill-behavior change, no third-party integration — it makes the existing suite green again on dev.

What alternatives did you consider?

1. Restore the pruned content to the reference files so the tests pass unchanged — rejected: it directly contradicts e7ddc25's stated intent, and the restore-direction attempt PR Document Antigravity tool mappings for skill loading #1902 ("Document Antigravity tool mappings for skill loading") was already closed. Reintroducing the boilerplate would re-break what the maintainer deliberately trimmed.
2. Delete the two content-assertion tests entirely — rejected: the reference files still exist and still document harness-specific mappings worth guarding. Trimming the assertions keeps meaningful coverage (subagent dispatch, task artifact, SKILL.md link) while dropping only the pruned tokens.

Does this PR contain multiple unrelated changes?

No. One root cause — e7ddc25 orphaned the per-harness mapping-content tests. Both files fail for the identical reason and are fixed the identical way. One problem.

Existing PRs

• I have reviewed all open AND closed PRs for duplicates or prior art
• Related PRs: Document Antigravity tool mappings for skill loading #1902 (closed) attempted the opposite fix — restoring the pruned antigravity skill-load docs; its closure is what confirms the reference should stay lean and the tests should adapt instead. writing-skills and porting guide still link claude-code-tools.md / copilot-tools.md, deleted in e7ddc25 #1969 / Fix dead references to pruned claude-code-tools.md/copilot-tools.md (#1969) #1970 address a different slice of e7ddc25 fallout (dead markdown links to the fully-deleted claude-code-tools.md / copilot-tools.md), not these test assertions. No open PR targets these two test files as a focused fix. (Several large, CONFLICTING, stale sweep-PRs incidentally include these files but do not repair the assertions.)

Environment tested

Harness	Harness version	Model	Model version/ID
Claude Code	2.1.143	Claude Opus 4.8 (1M context)	claude-opus-4-8[1m]

Tests run locally on macOS (bash + node --test).

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

N/A — this PR does not add a harness.

Evaluation

Not a behavior/skill change, so there is no eval-session delta to report. Verification instead:

• Before (on dev): tests/antigravity/test-antigravity-tools.sh → exit 1; tests/pi/test-pi-extension.mjs → 5 pass / 1 fail.
• After: antigravity → PASS; pi → 6 pass / 0 fail.
• Also ran tests/antigravity/run-tests.sh (all passed) and the repo shell-lint tests (pass). Confirmed each surviving asserted token is actually present in the pruned reference file.

Rigor

• If this is a skills change: … — N/A, not a skills change (tests only).
• This change was tested adversarially, not just the happy path — verified failing-first on both dev and main, and confirmed the new assertions match the pruned references.
• I did not modify carefully-tuned content (Red Flags table, rationalizations, "human partner" language) — no skill/prose content touched.

Human review

• A human has reviewed the COMPLETE proposed diff before submission — @gaurav0107 reviewed the full two-file diff and approved it.

[muunkky]

The antigravity half holds up. I stripped the mapping table out of antigravity-tools.md and your test correctly fails on invoke_subagent, which is the one token that only ever appears in the table. That's the right shape.

The pi half doesn't hold. All four of the new tokens also occur in the prose, so the assertion survives the thing it exists to detect. On your branch, with the pi mapping table deleted:

$ grep -v '^|' skills/using-superpowers/references/pi-tools.md > t && mv t skills/using-superpowers/references/pi-tools.md
$ node --test tests/pi/test-pi-extension.mjs
# pass 6
# fail 0

subagent matches the intro line, pi-subagents and Task both match the Subagents paragraph, and TODO.md matches the Task lists paragraph. Task is the one I'd look at twice: it matches "do not fabricate Task calls", a sentence whose entire job is to say Pi has no Task tool.

So the test passes on a pi-tools.md with no mapping table at all, which is the regression e7ddc25 caused in the first place.

Scoping the match to the table rows closes it:

const rows = text.split('\n').filter((line) => line.startsWith('|'));
assert.ok(rows.some((row) => /subagent/i.test(row)), 'mapping table documents subagent dispatch');
assert.ok(rows.some((row) => /todo|task/i.test(row)), 'mapping table documents task tracking');

That fails on the stripped file and passes on dev. Being straight about it: that snippet is the diff in my #1983, which does the pi half and nothing else. Take it or write your own — yours covers antigravity too, so it's the better PR either way and I'll close mine.

Disclosure: agent-assisted — Claude Opus 4.8 (claude-opus-4-8[1m]), Claude Code 2.1.208. Plugins: gitban (muunkky.github.io/gitban-site), frontend-design, skill-creator, claude-patent-creator-standalone. Not superpowers itself. Grounded in the runs above, against your branch at 8296a98.

[gaurav0107]

Thanks @muunkky — this is exactly right, and I reproduced it before changing anything. Stripping the table rows from pi-tools.md and re-running the test, my token list passed 6/0, so the assertion was surviving the regression it exists to catch. Task matching "do not fabricate Task calls" was the clearest tell.

Pushed the table-scoped version (0873545) — filtering to rows that start with |, then asserting /subagent/i and /todo|task/i against those rows. Verified it now fails on a table-stripped pi-tools.md and passes on dev. The approach and snippet are from your #1983 — credited in the commit message.

I checked the antigravity half against the same test: it already fails when the table is stripped (invoke_subagent only appears in the row), so I left it as-is. Since this PR now covers both halves, feel free to close #1983 — and thanks for doing the adversarial pass.

Disclosure: agent-assisted — Claude Opus 4.8 (claude-opus-4-8[1m]), Claude Code 2.1.143, reviewed and approved by @gaurav0107. Plugins unchanged from the PR description.

[cwest]

Backing this one — and I've closed my overlapping PR (#1992) in favor of it.

I hit the same red antigravity suite on dev and first went the other way:
restoring the pruned rows in antigravity-tools.md. Then I re-read #1847 —
that content was cut deliberately ("the harness-specific notes that still
carry weight") — so re-adding it fights the intent. Realigning the stale
tests, as you've done, is the right call.

Reproduced locally:

• tests/antigravity/run-tests.sh → green
• the pi mapping case flips red→green against the slimmed pi-tools.md
• a clean agy run with the plugin installed loads skills and auto-triggers
  brainstorming without those mapping lines present — the empirical version
  of "agents mostly don't need 'em"

Honest caveat: I couldn't run pi extension cases 2–5 headlessly (they load the
TS extension and my setup wouldn't), so I verified the mapping assertion, not
those four.

Thanks for taking the pi half down to the table rows too.