Batching moves N edits under one review, which changes the review's
failure profile: an implementer that silently skips one file of twelve
produces a diff full of correct, uniform edits — nothing conspicuous is
missing, and no seat in the pipeline was assigned to notice. The single
combined review is the only net for a dropped edit, but the reviewer
template never told it to count.
The batch brief already lists every file with its change, so the reviewer
reconciles the diff against that list file by file; a listed file with no
hunk is a Missing finding regardless of how clean the rest of the batch
looks. Conditional on a multi-file brief, so single-task reviews are
unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
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>
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.
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.
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.
Five blocks of benefits and cost/benefit selling aimed at a reader who
has already invoked the skill; the vs-Executing-Plans comparison also
duplicates the one under When to Use. Integration section untouched
(PR #1932 owns it).
The list-style Integration sections in subagent-driven-development and
executing-plans duplicated references that already exist where the flow
uses them (process digraph, When to Use, prompt templates, Step 3), so
they added maintenance cost without carrying behavior. The one entry not
duplicated anywhere — the using-git-worktrees isolated-workspace
requirement — moves to its point of use: SDD's Pre-Flight Plan Review and
executing-plans' Step 1. Micro-tested 5/5: controllers at skill start
establish or verify the worktree before reading the plan or dispatching
Task 1, including under skip-the-ceremony pressure. The prose Integration
sections in requesting-code-review and other skills are unchanged — they
carry placement content, not an index.
The per-worktree workspace test was added but never registered in
run-skill-tests.sh, so it only ran when invoked by hand. Add it to the
fast unit-test array alongside the other pure-shell test.
Also document, in the Durable Progress section, that the ledger now
lives in git-ignored working-tree scratch, so `git clean -fdx` deletes
it — recover from `git log` if that happens.
Reviewer tripwire (Calibration): a plan-mandated defect IS a finding,
reported as Important and labeled plan-mandated — the plan's authorship
does not grade its own work.
Controller rule (review loop): a plan-mandated finding, or any finding
conflicting with the plan's text, escalates to the human like any plan
contradiction — never dismissed because the plan mandates it.
E35 micro (frozen 0a98 replay, sonnet reviewer, 6v6): without the
tripwire 0/6 reports give the controller anything to escalate on (all
Approved, defect endorsed as spec-required); with it 6/6 report the
defect as a labeled finding.
E30 replay: the planted-DRY catch is causally determined by the
controller-composed constraints block (0/6 with process-shaped vs 5/6
with the spec's own wording). E31 micro: this recipe doubles the rate
at which composed blocks carry the spec's cross-component relationship
(6/6 vs 3/6). Affects dev and the redesign equally (E29: both 4/5).
scripts/review-package generates the reviewer's input deterministically:
commit list, stat summary, and net diff with -U10 context, written to a
file from an explicit BASE. Live runs showed controllers improvising
'git diff HEAD~1..HEAD', which silently truncates multi-commit tasks,
and svelte's five fix dispatches shipped without re-running any tests —
fix dispatches now explicitly carry the implementer's
re-run-and-report contract.
The skill read as a changelog: 'combined task review,' 'one reviewer,
one reading,' 'one dispatch,' and an example still showing diffs pasted
into prompts. A reader who never saw the two-reviewer design has no
referent for 'combined.' Prose now states the design directly, and the
flowchart/example reflect the diff-file handoff.
Paste adoption stayed at 0/15 even as a Red Flag — and the controller's
reluctance is locally rational: pasting loads the diff into the (most
expensive) controller context permanently, while a reviewer self-fetch
costs a few cheap turns. The diff-file handoff is cheap for both sides:
the controller redirects git diff to /tmp without reading it, and the
reviewer gets the whole change in one Read call.
Fourth planted-defect failure mode: the implementer's self-report said
'noted mild structural duplication; left unabstracted per YAGNI' and the
reviewer deferred to that framing, rating the duplication no finding at
all. The pre-judging keeps relocating — controller prompt, then reviewer
calibration, now the implementer's report. Rationales are claims; they
never downgrade severity.
Adoption was 6/11 reviews on fractals and 0/17 on svelte when phrased
as guidance; reviewers without the diff re-derive it by hand, which is
the single largest remaining reviewer cost. Now a Red Flags Never entry
and a REQUIRED marker on the template placeholder.
With merged review, a planted verbatim-duplication defect shipped: the
reviewer rated it Minor (YAGNI) under the strict cannot-be-trusted
definition of Important, and the Minor-rolls-up rule meant no fix was
ever dispatched and the final review never saw the finding. Calibration
now names merge-blocking maintainability damage (verbatim duplication,
swallowed errors, assertion-free tests) as Important, and controllers
must paste accumulated Minor findings into the final review dispatch.
Iteration-1 profiling: implementers and per-dispatch overhead dominate
(429 of 686 subagent turns; controller coordination is half the dollars
and scales with dispatch count), reviewers are individually lean, and
the controller pasted the diff in only 2 of 22 review dispatches when
the guidance was phrased as optional.
Changes: spec-reviewer-prompt.md + code-quality-reviewer-prompt.md
replaced by task-reviewer-prompt.md (one reviewer, one reading of a
pasted diff, two verdicts: spec compliance ✅/❌/⚠️ and task quality);
one fix dispatch can address both kinds of findings; controller now
runs git diff itself and pastes it (imperative, not optional);
implementers run focused tests while iterating and the full suite once
before committing; flowchart, example, Red Flags, tool tables updated.
The broad final whole-branch review is unchanged.
Round-2 fractals eval regressed to 70min/32.2M tokens (vs round-1's
42.8min/14.5M) while reaching baseline-parity quality. Per-subagent turn
profiling attributed it to: haiku dispatches taking 2-3x the turns of
sonnet (678 of 1197 subagent turns), reviewers re-fetching diffs by hand
(518 Bash calls), and evidence-rule narration. Changes: turn-count-beats-
token-price model guidance; controllers paste small diffs into reviewer
prompts (reviewers then need few or no tool calls); evidence scoped to
findings and would-be-bare-yes checks; Important defined as cannot-trust-
until-fixed with coverage suggestions Minor; fixes dispatched only for
Critical/Important.
Resumed the offending eval controller session and asked it why it
pre-judged despite the rule being in context. Its retrospective: the
motive was avoiding a review loop, the abstract rule was read but not
applied at the moment it governs, and a phrase-level trigger ('do not
flag', 'at most Minor', 'don't treat X as a defect', 'the plan chose')
would have fired where the principle did not.
Second observed instance: with the Constructing Reviewer Prompts rule
already live, a controller still wrote 'do not treat that duplication as
a defect to fix — the plan chose it; you may note it as a Minor
observation at most' into a quality reviewer dispatch, fabricating plan
intent from the plan's example snippet. Promote the rule to the Red
Flags Never list and name the rationalization.
Live eval deliverables shipped five polish defects; tracing each through
the transcripts showed three mechanisms, each now addressed:
- reviewers answered pointed checklist items with unsupported yes
(evidence rule: every What-to-Check answer needs file:line evidence)
- no reviewer ever saw the design's global constraints (controllers now
paste binding constraints into task requirements)
- test output noise was invisible everywhere (pristine-output checks in
implementer self-review and quality review)
In live eval runs, controllers given judgment-based model selection
stopped passing a model at all; the omitted parameter inherits the
session's top-tier model, silently making every subagent maximally
expensive (one run dispatched 26/26 reviewers on the session model).
A live eval run of sdd-quality-reviewer-catches-planted-defect caught the
SDD controller fabricating a plan constraint and instructing the quality
reviewer not to flag the planted DRY violation. The duplication shipped.
Constructing Reviewer Prompts now bans suppression directives alongside
open-ended broadening directives.
