Skip to content

sync: upstream pstack bf0c7cd (0.15.42) - #53

Merged
negoro26 merged 2 commits into
mainfrom
sync/upstream-bf0c7cd
Nov 16, 2026
Merged

negoro26 merged 2 commits into
mainfrom
sync/upstream-bf0c7cd

Conversation

@negoro26

@negoro26 negoro26 commented Nov 16, 2026 •

Copy link
Copy Markdown
Owner

Why

The scheduled upstream sync (run 43721586662) failed its pre-flight at bf0c7cd, so the port stayed on c742dd6 (0.15.36). Users of the omp port need upstream's 0.15.42 text with no Cursor mechanics, no model slugs, and nothing dropped. This PR moves the pin to bf0c7cd and gets check-port.sh to PASS there. Upstream collapsed how and why to one-file skills and moved every default model, so 23 rules went stale and 8 swallow spans changed.

Upstream commits since c742dd6

bf0c7cd	feat(pstack): new default models, one-file why and how, statechart fixes, sharpen and reflect text (#1236)

Upstream version moves 0.15.36 to 0.15.42, so RELEASE resets to 0 and the catalog publishes 0.15.4200.

Rule decisions

Rule numbers are rules.sed lines at this PR's head.

Rewritten

  • 84. Upstream now says "go to your hardest tasks model (default claude-fable-5-2-max)". The left side matches that, and the capability stays "your strongest judgment model". Rule 93 claims the line first, so 84 is a consumed net.
  • 93. The Defaults paragraph gained two sentences after "Prose and judgment read". The left side now ends at "rerun /setup-pstack." The replacement carries the new panel default as "Review and design panels default to three runners, each on a different configured model family when the roster supplies one."
  • 119. Upstream's how now spawns one explainer with subagent_type: generalPurpose and readonly: true. The rule that used to match an older form of this sentence now matches it. It starts one explainer in one task call, uses a discovered explainer or the default worker, and carries read-only posture in the brief. Its exemption row is gone.
  • 143. Architect's default runners changed from two slugs to three. The left side matches a three-slug list.
  • 144. Arena step 3 now defaults to three runners and parses families by longest prefix. The left side matches the new text. The replacement adds "Run three by default, and never fewer than two."
  • 145. The arena judge test moved from family to vendor, with claude-fable-* and claude-* as one vendor. The left side ends at "the vendor check did not run." The replacement uses vendor throughout and adds "Two model families from one vendor, such as two Claude lines, count as one vendor."
  • 148, 176. Swarm's worker default and the multi-phase plan's lane default moved from grok to claude-opus-5-5-high-fast. Both now match any single default slug, as rules 107 to 111 already do.

Added

  • 186. Recall step 4 now spawns its own investigators "in agent mode because readonly strips MCP access". The rule starts them in one task call and says each needs its source's MCP tools, which omp task agents inherit. The inheritance claim is the port's existing wording at rules.sed:20 and :562.
  • 201. Tier for Claude fast slugs such as claude-opus-5-5-high-fast, mapped to "your fast code model". Upstream uses it as the swarm workers default.
  • 203. Tier for Gemini slugs such as gemini-4-carbon-high, mapped to "your strongest model from another vendor". Upstream uses it as the third panel runner.

Both tiers are consumed at this pin because sentence rules claim every occurrence. They keep untiered slugs at "none" if a sentence rule misses.

Exempted, upstream removed the text at bf0c7cd

Each row uses the file's "upstream removed X at bf0c7cd; retain ... if that sentence returns" form. The repository's convention keeps such rules as nets, as it did for Orchestrate at c742dd6 and interrogate at ab1f702.

  • 19, 20. how's and why's spawn-contract paragraphs. Their swallow rows now hold the empty hash, so the text's return fails the ledger.
  • 22, 23. why's inline investigator and synthesizer model clauses.
  • 100. why's per-spawn model paragraph.
  • 122. how's explorer fan-out.
  • 168. The `model`: the <role> line token net. why carried the last instance.
  • 171, 172. why's investigator fan-out and synthesizer step.
  • 199. The grok fast-code slug tier. Upstream has no grok slug left.
  • 366. why's MCP discovery step.
  • 384. The how explorer and how explainer role labels.
  • 394, 395. how's simple path and explorer synthesis.
  • 447. how's explainer prompt, deleted with references/.
  • 558. why's vague-target step.

Exemption rows changed

  • Deleted. The row for rule 193 (claude-fable tier). That rule fires again, on claude-fable-5-2-*.
  • Deleted. The row for rule 119's old fragment, since the rule now matches upstream.
  • Re-reasoned. Rules 117, 120, 121, 123, 181, 401 and 446. Their rows said "a different rule now covers the current phrasing", and the covering rules are the ones exempted above. They now say upstream removed the text at bf0c7cd.

Swallow ledger reviewed

  • 19, 20. Empty hash, as above.
  • 36 (Match the model). Upstream added Cursor's slug ladder, namely high as a default level, max for hardest tasks, Fable as its own family, and the levels and fast variants Gemini and Fable 5.2 lack. omp's task item takes no model, so none of it has a counterpart. The replacement is unchanged.
  • 71 (poteto-help portability). Upstream dropped /why and narrowed per-role models to some skills. The replacement names no skill list.
  • 92 (poteto-agent rule). Upstream dropped how and why from the skills that set their own agent. The replacement names no skill list.
  • 93 (Defaults). Carried the three-runner panel default. Dropped the # defaults: 2 setup-version notice, because setup-pstack is excluded and omp has no rule file to version.
  • 144 (arena runners). Carried the three-runner default and the two-runner floor. Prefix parsing and slug fallback have no omp counterpart.
  • 145 (arena judge). Carried the vendor test and the vendor-check wording. Pool order by slug and the default-pool retry are Cursor slugs.

Other

  • slug-allowlist.txt loses its one entry. Upstream replaced the gpt-4 rename example in reflect/references/synthesizer.md with parseConfig, so the entry was stale. The catch-all comment no longer cites that example.
  • Deleted an orphan comment claiming the how skill branches on Step 2a and 2b.
  • owned drift passes. Neither owned path has an upstream counterpart, and neither names how's or why's old pipeline.

Content-loss result

The script below checks every sentence of every upstream file that changed between c742dd6 and bf0c7cd against the generated tree.

  • 31 changed files, 14 deleted upstream, 1 excluded (skills/setup-pstack).
  • 6379 sentences appear verbatim.
  • 315 sentences are a named rule's translation. 100 sit on lines new at bf0c7cd. 215 sit on unchanged lines that earlier syncs reviewed.
  • 0 sentences are neither.
  • None of the 14 upstream-deleted files is in the tree.

The 100 new-line translations map to rules 35, 36, 71, 92, 93, 107 to 111, 119, 139, 143, 144, 145, 148, 176, 186, 261, 431 and 463. Every sentence a wildcard drops is one listed under "Swallow ledger reviewed". The rest are the single-sentence translations described above, or existing rules that fire on the reworded lines unchanged.

Leftover Cursor terms in plugins/pstack/{skills,agents}. There are 0 hits each for subagent_type, generalPurpose, run_in_background, pstack-models.mdc, ~/.cursor, `Task`, Task tool|subagent|spawn|call|schema, cloud agent, AskQuestion, environment: "cloud", cursor-team-kit, and any claude-, gpt-, grok-, gemini-, opus- or composer- slug. readonly appears in Markdown only as "there is no task readonly field" (arena, how, reflect) and in pstack-omp's own field list. Every other hit is a TypeScript readonly modifier.

content-loss.py
#!/usr/bin/env python3
"""Every sentence of every upstream file that changed between two shas, checked against the
generated tree. A sentence is `verbatim` when it appears in the generated file unchanged,
`translated` when it does not but a rules.sed rule fires on its raw line (the rule numbers are
listed), and `lost` when neither holds. Usage: content-loss.py <repo> <canon> <old> <new>."""
import re, subprocess, sys
from collections import defaultdict
from pathlib import Path

repo, canon, old, new = sys.argv[1:5]
port = Path(repo) / "omp-port"
marked = subprocess.run(["bash", "-c", f". {port}/lib.sh && mark_rules"], capture_output=True, text=True, check=True).stdout
Path("/tmp/sync-bf0/marked.sed").write_text(marked)
excluded = [l.split("#")[0].strip() for l in (port / "excluded.txt").read_text().splitlines()]
excluded = [e for e in excluded if e]

def git(*a):
    return subprocess.run(["git", "-C", canon, *a], capture_output=True, text=True, check=True).stdout

def fired(line):
    out = subprocess.run(["sed", "-E", "-f", "/tmp/sync-bf0/marked.sed"], input=line + "\n", capture_output=True, text=True).stdout
    return sorted({int(m) for m in re.findall(r"@([0-9]+)@", out)})

def translate(text, rules=str(port / "rules.sed")):
    return subprocess.run(["sed", "-E", "-f", rules], input=text + "\n", capture_output=True, text=True).stdout.rstrip("\n")

rules_sed = (port / "rules.sed").read_text().splitlines()
ranged = [n for n, r in enumerate(rules_sed, 1) if r.startswith(("/", "\\%"))]

def ranged_fired(raw, line):
    hits = []
    for n in ranged:
        Path("/tmp/sync-bf0/one.sed").write_text(rules_sed[n - 1] + "\n")
        out = set(translate(raw, "/tmp/sync-bf0/one.sed").splitlines())
        if line not in out:
            hits.append(n)
    return hits

def sentences(line):
    return [s.strip() for s in re.split(r"(?<=[.!?:])\s+", line) if len(s.strip()) > 3]

stat = defaultdict(int)
by_rule = defaultdict(list)
lost = []
deleted = []
for row in git("diff", "--name-status", old, new, "--", "pstack/skills", "pstack/agents").splitlines():
    st, path = row.split("\t")[0], row.split("\t")[-1]
    rel = path.removeprefix("pstack/")
    if any(rel == e or rel.startswith(e + "/") for e in excluded):
        stat["excluded files"] += 1
        continue
    gen = Path(repo) / "plugins/pstack" / rel
    if st == "D":
        deleted.append(rel)
        if gen.exists():
            lost.append((rel, "(upstream deleted it, the tree still has it)", []))
        continue
    raw = git("show", f"{new}:{path}")
    before = set(git("show", f"{old}:{path}").splitlines()) if st == "M" else set()
    text = gen.read_text() if gen.exists() else ""
    stat["files"] += 1
    for line in raw.splitlines():
        for s in sentences(line):
            if s in text:
                stat["verbatim"] += 1
                continue
            t = translate(s)
            if t != s and t.strip() in text:
                rules = fired(s)
            else:
                rules = fired(line) or ranged_fired(raw, line)
            if rules:
                stat["translated"] += 1
                by_rule[tuple(rules)].append((rel, ("" if line in before else "[new] ") + s))
                stat["translated, new at target" if line not in before else "translated, unchanged line"] += 1
            else:
                stat["lost"] += 1
                lost.append((rel, s, []))

print("changed upstream files:", dict(stat), "deleted upstream:", len(deleted))
for d in deleted:
    print("  deleted upstream:", d)

for rules, items in sorted(by_rule.items()):
    print(f"\n== translated by rules.sed:{','.join(map(str, rules))}")
    for n in rules:
        print(f"   rule {n}: {rules_sed[n - 1][:160]}")
    for rel, s in items:
        print(f"   {rel}: {s}")
print(f"\n== lost: {len(lost)}")
for rel, s, _ in lost:
    print(f"   {rel}: {s}")

Run as python3 content-loss.py "$PWD" /tmp/cursor-plugins c742dd61706fbacc1b95f1a20b7b988cf096ce6c bf0c7cd80abdbf7fe47b634978a7d1b4071f71a7 from the repository root. It expects /tmp/sync-bf0/ to exist for scratch files.

Verification

All commands ran in this branch's worktree at 5bcc994 with /tmp/cursor-plugins holding both shas.

  • CANON=/tmp/cursor-plugins bash omp-port/sync-upstream.sh bf0c7cd80abdbf7fe47b634978a7d1b4071f71a7 printed pin c742dd6 -> bf0c7cd and version 0.15.4200, upstream 0.15.42 at port revision 0.
  • CANON=/tmp/cursor-plugins bash omp-port/check-port.sh printed check-port: PASS. 47 checks report 44 PASS, 2 REPORT and 1 SKIP, since the target is the pin. untiered slugs reports none. rule liveness reports 134 rules that fire nothing, namely 101 stale and exempted, 32 consumed, and the catch-all.
  • bash omp-port/test-gate-mutations.sh printed mutation tests: PASS.
  • bash omp-port/test-sync-owned-paths.sh printed sync owned paths: PASS.
  • bash omp-port/test-upstream-sync.sh printed PASS upstream-sync workflow steps and check.
  • OMP_SMOKE_REQUIRED=1 bun omp-port/smoke-plugin.mjs passed on omp 18.20.3 through -e, omp plugin link, and the marketplace install, each with 59/59 skills registered.

Review round

  • Claude reviewer: APPROVE (0.88). It rebuilt from the old pin and got an identical tree. 0 unexplained sentences out of 5003. It confirmed the three-runner default, the two-runner floor and the vendor judge against upstream text.
  • Qwen cross-family reviewer: VERDICT: PASS.
  • Follow-up commit: moved the Claude fast-slug tier (claude-[a-z]+-[0-9.-]+-([a-z]+-)?fast) above the fable and opus judgment tiers. Before, an effort-less claude-opus-5-5-fast became "your strongest judgment model". It now maps to "your fast code model". No generated file changes at bf0c7cd. This renumbers the fable tier to 193 and the grok tier to 199. Rule numbers above refer to the new head.

- Move the pin from c742dd6 (0.15.36) to bf0c7cd (0.15.42). RELEASE resets to 0, so the catalog publishes 0.15.4200.
- Rewrite rules 83, 92, 142-144, 147 and 175 for upstream's new default models and the arena vendor test, and rule 118 for the one-explainer how.
- Add a recall investigator rule and Claude-fast and Gemini slug tiers.
- Exempt 16 rules whose text upstream removed with the multi-agent how and why, and re-reason 7 exemptions that pointed at those rules.
- Record the new swallow spans for rules 19, 20, 36, 71, 91, 92, 143 and 144, carrying the three-runner panel default and the vendor judge into the replacements.
- Drop the stale gpt-4 allowlist entry; upstream replaced that example.
@negoro26
negoro26 merged commit 1c91fef into main Nov 16, 2026
3 checks passed
@negoro26
negoro26 deleted the sync/upstream-bf0c7cd branch November 16, 2026 10:51
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants