Repository navigation
Bump skills-vendor pin: init-socraticode's #115 policy split is not yet vendored #41
Description
skills-vendor/gregoryfoster-skills is pinned at 3fc7b71; upstream main is c3f40f7.
#35 brought this repo's ## Code Exploration Policy block onto the shape init-socraticode adopted in gregoryfoster/skills#115 (upstream commit 9125148) — Variant A in AGENTS.md, the full reference in docs/SOCRATICODE.md, and repo-authored content rescued into ## Code Exploration Notes (repo-specific) outside the marker pair. That saved 489 tokens on a section curating-context is contractually forbidden to touch.
The vendored skill does not know about any of it. Verified against the pin:
vendored 3fc7b71 |
upstream 9125148 |
|
|---|---|---|
mentions docs/SOCRATICODE.md |
0 | yes |
| tool-table rows in the block | 8 | 3 |
| step 1a (rescue repo-authored content) | absent | present |
So an init-socraticode re-run today would:
- take the marker-pair branch and restore the pre-split eight-row block into
AGENTS.md, reverting the saving and pushing the file back toward budget; - never write or overwrite
docs/SOCRATICODE.md, leaving it an orphan duplicating content that just came back inline.
## Code Exploration Notes (repo-specific) is safe either way — the old sweep targets only ## Code Exploration Policy, and the rescue heading is deliberately different.
Ask
Bump the submodule and re-run init-socraticode, reviewing its rescue report. Deliberately not bundled into #35: that PR is a documentation curation, and a pin bump pulls in #115 plus unrelated upstream work (#107 graph-yield gating, #141 self-budget gate, #144 batch clearance) that deserves its own review.
Noted in docs/SOCRATICODE.md's header so anyone re-running before the bump knows what they will get.
Activity
#35 docs: apply CR round-2 findings (#8-#11)
gregoryfoster commentedon Aug 15, 2026
Done in #43 (merged as
0257f3c).The pin bump came from #39's hook, not by hand — registering the
SessionStartentry made the refresh hook fire onmain, and its own auto-commit (36326c4) movedskills-vendor/gregoryfoster-skills3fc7b71→2e1cf28. Upstreammainhad moved past thec3f40f7quoted here; the bump is 153 commits.obra-superpowerscame along to v6.3.0.Rescue report: nothing moved
The destructive re-run predicted here does not materialize, and the reason is the one this issue identified as the safe case. #35 had already landed the marker pair, so Phase 3 takes the replace-between-markers branch — and
## Code Exploration Notes (repo-specific)sits outside that pair, so it was never inside the replaced span. The whole-span replace only exists on the unmarked branch.The re-run against the new skill also writes
docs/SOCRATICODE.mdrather than orphaning it, so both failure modes listed here are closed.What actually changed
AGENTS.mdpolicy blockok); one line changed — the prefetch now names.claude/hooks/socraticode-reminder.sh. Block unchanged at 443 tokens, so #35's saving is intactdocs/SOCRATICODE.md.claude/hooks/socraticode-reminder.sh.claude/hooks/socraticode-health.shdocs/SKILLS.mdThe unrelated upstream work this issue flagged as deserving its own review — #107 graph-yield gating, #141 self-budget gate, #144 batch clearance — is reviewed in #43's description.
Verification
settings.jsonmerge idempotent — a second pass is byte-identical## Code Exploration Notesrule —codebase_context_searchreturns the newdocs/SKILLS.mdtextFiled upstream
gregoryfoster/skills#177.
mcp-driver.mjsguards dispatch onpath.resolve(argv[1]) === fileURLToPath(import.meta.url), which never matches through a symlink — so invoked at the path SKILL.md documents (skills/init-socraticode/scripts/…, amanaging-skillssymlink) it exits 0 having printed nothing. Worse,socraticode-health.shresolves that symlink ahead of the realskills-vendor/path, so the health hook cannot report on any repo in the cohort — the never-measured green light #107 exists to prevent. Worked around here viaSOCRATICODE_DRIVER; documented indocs/SOCRATICODE.md.chore: calibrate the context estimator per file
gregoryfoster commentedon Aug 16, 2026
Follow-up: two code-review rounds after the close corrected three things in the re-run's output.
#177workaround was machine-local.SOCRATICODE_DRIVERsat only in gitignored.claude/settings.local.json, so every other checkout shipped a health hook that resolved the brokenskills/symlink and reported clean without measuring anything. The value is a repo-relative path, so it moved to the committed.claude/settings.jsonenvblock — verified reaching spawned processes, with the hook now reporting from committed config alone (docs+test: apply CR round-1 findings (#1-#8) #45).#177note went intodocs/SOCRATICODE.md, six lines under its own "a re-run overwrites this file wholesale" header — this issue's own concern, one level down. Moved toAGENTS.md→## Code Exploration Notes (repo-specific), which survives a re-run. The write guard caught the first draft 68 tokens over budget, so the detail went todocs/SKILLS.mdand only the rule stayed inline (docs+test: apply CR round-1 findings (#1-#8) #45).node skills/init-socraticode/scripts/mcp-driver.mjs— the invocation that exits 0 printing nothing.docs/SKILLS.mdnow carries an after-a-re-run checklist; two of the three items fail a test, and the checklist covers the one that cannot (docs+test: apply CR round-2 findings (#9, #11, #12) #46).Also filed upstream from this work: skills#177 (driver no-ops through the vendoring symlink, and the health hook resolves that symlink first), skills#179 (copy-vs-symlink contradiction), skills#180 (health hook measures
., so a worktree session reports a healthy index as broken).Shipped:
c5698e9,86ceda6,b393590,c29cbe6(PRs #43, #45, #46). Context artifacts re-indexed after every doc change, per this repo's own rule.