Document Pi mappings for legacy skill vocabulary#1903
Conversation
|
Hi @tianma-if — I'm Claude, an AI agent (Fable 5, running in Claude Code), posting from Jesse (@obra)'s account at his direction. Jesse had me run a skeptical triage of all 340 open superpowers items — every factual claim tested against the current Thanks for opening this. The technical claim checks out: That said, I can't evaluate this PR in isolation: it's #1903 of ten (#1901-#1910) you opened between 04:19:42 and 04:20:16 UTC on 2026-07-03, one every 3-4 seconds, spanning unrelated subsystems, each with the identical "Human partner who reviewed this diff: msh01" claim and the same "user asked to find real, high-quality contribution candidates" framing. No one reviewed ten independent cross-cutting diffs in 34 seconds. Our CLAUDE.md is explicit that batches like this get closed regardless of individual merit - please resubmit as a single PR with a genuine human review pass. Separately, even on the merits: this restores the exact Read/Write/Edit/Bash/Skill table rows that commit e7ddc25 ("Prune per-harness tool-mapping boilerplate") deliberately removed nine days earlier, reasoning they "restate guidance modern agents already follow." A trim of the test's assertions (or a one-sentence legacy-vocabulary note) would fix the real breakage without reverting that decision. If any of the evidence above is wrong, reply here — Jesse reads these, and closures can be revisited. |
|
This test is still red on The cause is the one @obra identified here: @tianma-if — this is your find, and you get first refusal. @obra's note here spells out the fix he'd take, and it is not the patch you submitted: a trim of the test's stale assertions, rather than restoring the rows If you'd rather not, or don't reply, I'll open it as a single standalone PR against Scope, if it lands on me: Verification is the whole story and it's mechanical: red → green, Disclosure: agent-assisted — Claude Opus 4.8 ( |
|
Opened the trim as #1983 (draft, against It's one line: the assertion list now checks the mappings that survived Disclosure: agent-assisted — Claude Opus 4.8 ( |
'One PR at a time' was the wrong lesson. What upstream polices is 'an agent pointed at the issue list and told to fix things' — no understanding, no prior engagement, no human in the loop. obra closed obra#1903 for ten PRs in 34 seconds with an identical fabricated human-reviewer claim; the diffs were not the problem, the absence of work behind them was. So the gate is not 'how many' but 'can you point at where you socialized it'. A comment on the issue thread, posted before any code, asking for a sanity check — that is the alibi a trawler cannot fake after the fact. Stage 2 is not etiquette, it is evidence. Adds the spray-vs-genuine table, and puts the gate in the gitban-pr overlay, which had no guard at all despite being the skill that opens the PRs.
Fair criticism: I could not explain consistently what obra wants, and kept contradicting myself across turns — first that our process-to-output ratio was the over-engineering he punishes, then the reverse. The cause is structural. This file was a list of eight symptoms with no underlying model, so anyone using it (me) reasons case-by-case and lands wherever the last example pointed. One disease explains all eight. He is a maintainer buried under machine-generated submissions that look right and aren't, so he is not reviewing the code — he is running a cheap falsification test on one question: did a mind make contact with reality here? From that, without special pleading: - correct code doesn't save you (obra#1904 was 'cleanly mergeable' and closed) - every sentence is a claim and he RUNS it (obra#1797, obra#1906, obra#1925, obra#1166) - depth is demanded, complexity is punished — different axes (obra#1797 vs obra#668) - the artifact of thought is not evidence of thought; ship the residue he can re-run, not the deliberation (0 of 42 merges linked a PRD) - the code is that way on purpose (obra#1168, obra#1903, obra#1882) - venue first (10 of 42), not-our-defect (6 of 42) - the trawl is the crime, not the count (measured: 3.7% vs 3.2%) - silence is the baseline, not a verdict The anti-patterns stay — they carry the receipts. But they are now consequences of a stated model rather than eight things to memorise.
3 participants
Document Pi mappings for legacy skill vocabulary
Base:
obra/superpowers:devHead:
msh01/superpowers:codex/document-pi-tool-mappingsWho is submitting this PR? (required)
What problem are you trying to solve?
node tests/pi/test-pi-extension.mjsfails onupstream/devbecauseskills/using-superpowers/references/pi-tools.mddoes not document the legacy skill vocabulary the test expects (Skill,Task,TodoWrite, read/write/edit/bash mappings).The failing assertion was:
What does this PR change?
Adds a compact Pi mapping table for legacy Claude-style tool names and a short "Legacy skill vocabulary" section that maps those names to Pi file/shell/subagent/task-list capabilities.
Is this change appropriate for the core library?
Yes. Pi is already declared in
package.jsonand tested intests/pi/; this updates the existing Pi reference to satisfy the existing platform test.What alternatives did you consider?
I considered weakening the Pi test, but the test is guarding useful platform-adaptation documentation. The smaller fix is to document the expected vocabulary in the existing Pi reference.
Does this PR contain multiple unrelated changes?
No. It only updates
pi-tools.md.Existing PRs
Environment tested
New harness support (required if this PR adds a new harness)
Not applicable. Pi support already exists.
Evaluation
node tests/pi/test-pi-extension.mjsfailed on missing mapping vocabulary; after, it passed.Verification run:
Rigor
Human review