Both PRs rewrote the same preflight paragraph. Resolution is the composed
text published in #2080's description — the configuration the 3/3+3/3
composed eval grades ran.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The breaker's two ledger formats wrote lowercase 'ruling' (parked
findings, load-bearing adjudications), so the Finish section's
collect-every-`Ruling:`-line step missed exactly the rulings made under
the most pressure. Field evidence from an independent eval rep: a
breaker-cap run adjudicated correctly, wrote everything to the
plan-scoped ledger, deleted the workspace at finish, and left no durable
trace of the adjudication.
Capitalize the two breaker formats to the canonical token, and make the
finish roll-up explicitly exhaustive across preflight, parked, and
breaker rulings.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
In controlled evals, an identical seeded-incoherence plan yielded 0-1/5
correct conflict resolutions when executed specless (controllers ruled
the conflicts 'internally explained') and 4-5/5 with the spec merely
present and named — even with no other skill-text changes. Cross-task
coherence turns out to be adjudicable only against ground truth above
the plan; this change makes that ground truth travel with the plan.
Claude-Session: https://claude.ai/code/session_0185AJr98gHx5EmwqNeft4Sy
The pre-Task-1 conflict scan currently permits 'the scan is clean' with
no evidence the scan happened — mined sessions show controllers skipping
straight to dispatch and plan conflicts surfacing mid-execution as
blocking questions. Requiring the scan to emit one row per task pair
sharing a file/interface and one row per task's self-consistency turns
the claim into an artifact; in controlled evals the table appeared 3/3
with conflicts surfaced pre-dispatch, and the mechanism held 3/3 when
composed with the never-stall ruling change (#2077).
Claude-Session: https://claude.ai/code/session_0185AJr98gHx5EmwqNeft4Sy
Plans sometimes enumerate many tiny, same-shape edits (one-line fixes,
constant changes, a field added across files) as separate tasks. The
current loop dispatches a fresh implementer plus review per task, so a
12-micro-task plan costs ~24 subagent seats for what one subagent could
do in a single pass. In controlled evals on a micro-task plan, batching
cut cost 73% and dispatches 87% with better completion than control; on
a 5-non-trivial-task plan the rule correctly never batched (dispatch
counts and completion identical to control).
Claude-Session: https://claude.ai/code/session_0185AJr98gHx5EmwqNeft4Sy
A donated session sat dormant 8h48m waiting for a plan-conflict answer
that cost ~zero tokens to decide. Wrong-ruling rework is bounded;
stalls are not. This encodes the never-stall doctrine: plan conflicts,
ambiguities, and cap exceptions get a controller ruling recorded in
the ledger and work proceeds; only irreversible/destructive actions,
security-sensitive actions, out-of-worktree side effects (merge/push/
publish), and totally-broken plans remain hard stops. Rulings surface
in the Finish report instead of as mid-run questions.
Evals: 3/3 no-stall vs control 3/3 stall-at-preflight on a
seeded-conflict SDD plan; catastrophic guard 5/5 (every rep reaching a
seeded DROP TABLE step refused it); re-validated 3/3 after rebase onto
the current fix-PR text; composes cleanly with the evidence-bearing
preflight treatment.
Claude-Session: https://claude.ai/code/session_0185AJr98gHx5EmwqNeft4Sy
Round 2 proved the long-wait mechanism (65.1%->0.0% timeouts) but
20-38 min silent waits starved graders and let 1/51 children vanish;
bounded 5-10 min stretches with a status line and list_agents
reconcile keep the efficiency and restore observability.
Docs-only wait guidance in the platform reference changed nothing
(65.1% vs 67.1% baseline wait-timeout rate); the discipline now lives
in the controller loop the session actually re-reads.
60-78% of wait_agent calls timed out across every measured corpus;
waits are event subscriptions, so one long wait replaces dozens of
polls at identical wake latency.
Triggering battery: Claude Code classified a brand-new project bounded
3/3 by reading 'existing, understood flow' as genre familiarity — once
while explicitly noting the repo was empty. Gemini routed the same
prompt architectural 3/3.
The first fix-cycle battery moved the depth-2 leak from implementers
(9/9 baseline -> 0/6) to a final reviewer that spawned two
sub-reviewers; the contract now reaches every dispatched role.
Five claims contradicted by the Codex CLI source (V2 has no
close_agent; followup_task always reaches a child; role files attach
via agent_type; full-history forks accept model/effort; V2 spawn
allowlist). Citations: superpowers-autoresearch
docs/2026-07-29-codex-multiagent-v2-capabilities.md.
Live ceremony battery: bounded reps produced zero doc ritual (the
measured win) but 2/3 implemented before any approval turn; the
bounded path now states the stop explicitly.
Spike / bounded / architectural classification said out loud, one-way
upgrade ratchet, approval gate on every path. The measured pathology:
the absolute hard-gate wording forced bounded tasks into the full
two-document ritual 5/5 while a no-guidance control differentiated
paths natively.
Converges dev with the rebased release SHAs on main immediately, while
the trees are identical, so the merge is conflict-free and the next
dev -> main release shows only genuinely-new work.
Release notes for everything on dev since v6.1.1, plus the version bump
to 6.2.0 across all seven declared manifest files (bump-version.sh,
audit clean). Tagging and marketplace publication happen after the
dev -> main merge.
Follow-up to #2011 (which fixed the ./-prefix mismatch for the documented
pattern form): strip a leading ./ from the caller's pattern instead of
double-prefixing it into a never-matching ././ form, and also match the
pattern with '**/' collapsed, since find -path cannot match '**/' against
zero directory levels and silently skipped files directly under the base
directory (src/top.test.ts vs src/**/*.test.ts).
Adds a deterministic test suite for the script with a stubbed npm.
The prune in e7ddc25e removed the `## Subagent support` section from
antigravity-tools.md but left the inline cross-reference to it in the
dispatch table, so `[Subagent support](#subagent-support)` resolves to
nothing. An agent following the pointer to learn the difference between
the `self` and `research` subagent types lands nowhere.
Drop the dangling parenthetical. The guidance it pointed at survives in
the same table cell -- `self` for full-capability work, `research` for
read-only -- so no content is lost and the row still answers the
question the removed section answered.
gemini-tools.md carries the same cross-reference but retains its
`## Subagent support` heading, so its link is valid and is left alone.
Eight tasks across two repos: new re-review template, template/reference
alignment, full SKILL.md lifecycle restructure with move map, two
seeded-ledger fixture helpers, three quorum scenarios, and the RED/GREEN/
regression live-run campaign.
Review-fix loop gets resume-the-implementer semantics, scoped
re-reviews, a five-round circuit breaker, and controller adjudication
at trip. SKILL.md reorganizes by lifecycle; Red Flags converts to a
rationalization table. Brainstormed with Jesse 2026-07-15.
tests/claude-code/test-subagent-driven-development.sh failed
intermittently for two independent reasons:
- Budget mismatch: the file runs 9 prompts with a 90s timeout each
(810s worst case) inside the runner's 600s per-file ceiling, so slow
backend days produced spurious timeouts. Raise the runner default to
900s and fix the help text, which claimed the default was 300.
- Case-sensitive prose matching: the assert helpers grepped free-form
model output case-sensitively, but models capitalize the skill's own
headings — observed failures include "Do Not Trust the Report"
missing pattern "not trust" and a structured answer missing
"First:.*spec.*compliance". Match case-insensitively in
assert_contains/assert_not_contains/assert_count/assert_order, widen
two Test 5 keyword patterns to phrasings observed in real runs, and
make assert_order dump the output on failure the way assert_contains
already does, so the next flake is diagnosable.
Observed 3 failures across 4 runs before the change (timeout, two
distinct pattern misses); 3/3 consecutive full runs pass after it.
The packaging pipeline only worked on a Mac with default umask, for
three stacked reasons:
- The deterministic-metadata tar flags (--uid/--gid/--uname/--gname)
are bsdtar spellings; GNU tar rejects them, so the tar.gz archive
step died on Linux. Detect the tar flavor and use --owner=:0
--group=:0 --numeric-owner on GNU tar, which writes byte-identical
ustar headers (uid/gid 0, empty uname/gname).
- Staged file modes depended on two umasks canceling out: git archive
masks entry modes with tar.umask (git default 0002 -> 775), and the
unflagged tar extraction re-masked with the process umask (022 on
macOS -> 755, but 002 elsewhere -> 775). Pin tar.umask=0022 on the
archive call and extract with -p so staged modes are canonical
755/644 on every machine.
- The test's timestamp assertion parsed bsdtar's -tv column layout and
expected epoch 0 rendered in a US timezone ("Dec 31 1969"); GNU tar
uses different columns and UTC hosts render "1970-01-01". Assert
mtime == 0 via python3 tarfile instead, matching how the test
already checks zip timestamps.
tests/codex/test-package-codex-plugin.sh now passes on Linux/GNU tar;
the bsdtar branch preserves the exact flags that passed on macOS.
The SessionStart command string starts with a quoted path, which breaks
both Windows shells Claude Code may hand it to: PowerShell parses the
leading quoted string as an expression and dies on the next bareword
('Unexpected token session-start', #1751), and cmd.exe's /c quote rule
drops the outer quotes when the path contains a metacharacter, so a
profile dir like C:\Users\Name(External) truncates the command at the
'(' (#1918). Either way the bootstrap silently never loads.
Declare shell: "bash" on the hook. Claude Code >= 2.1.81 then resolves
Git for Windows and runs the polyglot's bash path directly — the same
route it already picks when it detects Git Bash — and when Git Bash is
missing it surfaces an actionable install prompt instead of a parser
error. Older versions ignore the unknown key and behave exactly as
before (verified live on 2.0.77 and 2.1.80).
Verified end-to-end with real claude sessions: Linux (hook fires,
bootstrap injected), Windows 11 + Git Bash under a path containing
'(' and a space (fires, 3276-char context), and Windows 11 without
Git Bash (actionable error replaces the #1751 ParserError, reproduced
verbatim as control).
Fixes#1751Fixes#1918
The start-of-skill ledger check is now scoped to the plan's own
workspace and keyed to the ledger's first line. Baseline eval (25/25
reps) showed controllers already refuse foreign ledgers — at a cost of
6-13 tool calls of cross-plan forensics per resume; plan-scoping makes
the answer structural instead. The workspace is deleted once the final
review is clean — git history is the durable record.
sdd-workspace now requires the plan file and resolves
.superpowers/sdd/<plan-basename>/; task-brief and review-package write
into their plan's directory (review-package gains PLAN_FILE as its first
argument). Follow-up plans in the same working tree can no longer collide
with a previous plan's briefs, reports, or ledger.
25/25 baseline reps refused the stale foreign ledger via git forensics;
the spec's evaluation section now states the honest claims: structural
fix + measured disambiguation-cost delta + same-plan-resume regression
gate, shipping with explicit maintainer sign-off in place of a failing
S1 baseline.
Three RED rounds (25 reps, three framings incl. faithful compaction
resume) never reproduced blind stale-ledger adoption: sonnet controllers
forensically refuse foreign ledgers, spending 6-13 tool calls per resume
doing it. Jesse approved shipping the full change with the eval re-scoped
to what is true: Task 1 compiles the existing RED evidence, Task 4 runs
GREEN on a truthful v3 fixture (real implementations, rotating authors)
with an S2 released-text control, measuring regression safety and the
disambiguation-cost delta instead of an error rate.
Fixture v1 tripped the Task 1 STOP gate for the right reason: its
ledgers cited fabricated hashes, so RED agents dismissed them via git
forensics (S1 passed for the wrong mechanism, the S2 resume control
failed 5/5). v2 executes plan A's tasks as real commits, gives both
plans five tasks so numbering is ambiguous, adds a symmetric
resume-uncertainty line to the scenario prompt, hard-stops if the S2
control fails twice, and drops rm -rf from cleanup (hook-gated here).
Five tasks: RED baseline eval (writing-skills Iron Law — before any
skill edit), plan-scoped scripts via TDD, SKILL.md durable-progress
rewrite with mismatch guard and end-of-plan cleanup, GREEN eval with
refinement loop, consistency sweep. Eval = 5 fresh sonnet subagents per
scenario per arm, hand-scored.
The .superpowers/sdd workspace has no plan identity and no end-of-life:
follow-up plans in the same worktree read the previous plan's ledger as
their own progress, and artifacts leak into git (observed in serf, three
contamination rounds and ad-hoc progress-p2/p3 workarounds). Structural
fix: per-plan workspace subdirs, ledger names its plan, delete the
workspace when the final review is clean.
e7ddc25 deleted claude-code-tools.md and copilot-tools.md but left
writing-skills and the porting guide's reference-integration table
pointing at them. State the current architecture instead: Claude Code's
personal-skills path inline, and "no adapter file needed" for the
harnesses that ride the Claude Code-compatible tool surface.
Reported by @rasibintang (#1969, with a fix proposed in #1970).
Fixes#1969
Drew Ritter