Anthropic documents that Claude Opus 5 verifies its own work, and that telling
it to double-check or to delegate verification to a subagent causes
over-verification -- token cost with no quality gain. The general subtraction
detector (BP-SUB-001) already surfaces those blocks for every user, with no
model-awareness at all.
`optimize --subtract --for-model <name>` adds the missing half. It ANNOTATES a
subset of the candidates --subtract already produced; it is not a second
detector and can never widen the candidate set. A second SUBTRACT_DETECTORS
entry would have collided with BP-SUB-001 on de-dup, and a prose-only signal in
the agent prompt would have been untestable.
There is no auto-detection, by measurement rather than omission: a CLAUDE.md has
no frontmatter and no resolvable target model, and this operator's own `route`
skill deliberately runs a different model per session -- the same file is read
by whichever model comes next. So the model is named, and the citation is
reported as conditional everywhere a human sees it (agent report copy, and the
Step 7a listing that is the last surface before an approval file).
Precision comes from the TARGET, not the verb list. Measured across the
409-file corpus: 392 BP-SUB-001 candidates, 31 (7.9%) carry a verify verb, and
0 also carry a reflexive or delegated target. Two independent raw-text greps
found 0 as well, so the zero is the corpus rather than an over-narrow regex.
Those 31 verb-only blocks -- "sjekk relevante config-filer", "Type-sjekk:
pyright", "To verify plugin functionality" -- are exactly the false positives a
verb-only version would have produced, which is BP-JUDG-001's 7/7 failure
arriving one lens over. The numbers live in the register entry's note and are
pinned by a test, because a session that cannot see the measurement reads the
zero as a broken detector and loosens it.
`recognized` is reported separately from `matchedCount`: a typo'd model name and
a genuinely clean config both yield zero, and without the distinction the CLI
would report a silent no-op as good news. Dogfooded on the real machine --
`opus-5` gives recognized:true/matchedCount:0, `oppus5` gives recognized:false.
source.published is absent because the guide carries no visible publish date;
its absence is asserted so a later session does not invent one to match the
other entries' shape. Both quoted sentences were verified verbatim 2026-08-12.
The payload stays additive -- forModel and per-candidate modelScope appear only
under the flag, so a plain --subtract run is byte-identical to before (asserted
on the serialized bytes, since a key set to undefined passes a shallow check).
Suite 1724 -> 1752 (+28). The one remaining failure is the pre-existing
drift-cli --output-file crash, untouched by this work.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BRuXt6tZyowi8QYNKLSHQm
Dogfooding `plan` + `implement` against a throwaway config surfaced one root
defect with many arms: the command templates treat consecutive fenced blocks as
one shell. They are not. Every ```bash fence runs as its own Bash call in its own
process, so a variable set in one block is empty in the next, and `$$` is a
different PID (measured: 21710 vs 22109).
The planner agent confirmed the sharpest arm at runtime, reporting that
`Mode: $RAW_FLAG` "arrived literally unsubstituted" — `--raw` was documented in
three command files while being functionally dead. A machine sweep found the same
root in 20 places across 9 files, well past the two the written fasit predicted:
- `$RAW_FLAG` read from non-shell agent prompts (analyze, plan, implement)
- `$TMPFILE` read across blocks (tokens, manifest, whats-active,
plugin-health) — each command could not read the file it had just written
- `$GLOBAL_FLAG` across blocks (fix)
- `$TODAY` never assigned in any block (campaign), passing
`--reference-date ""` to a write CLI in six places
- three `$$` temp paths handed to the Read tool (fix), which expands neither
All now follow the hardened drift.md pattern: a fixed literal path, or a
re-derivation inside each block that needs it.
Also fixed, all confirmed against ground truth rather than inferred:
- `implement` printed a rollback ID it never captured (the timestamp lived only
inside a command substitution) — the one message a user reads after a bad run
- `plan` reported "No analysis results found" for valid sessions, because Read
was pointed at a glob it cannot expand; now uses Glob and verifies the
analysis report exists before spawning the agent
- five phase commands wrote state.yaml with two of four required fields; since
the agent writes all four, a follow-up write silently deleted the rest
- `implement` promised rollback deletes created files; rollback deliberately
leaves them (M-BUG-26 still open) — the doc, not the engine, was wrong
- `implement` claimed a score delta with no pre-change measurement
- `verifier-agent` was told to write a report it has no tool to write
- dead `Task` tool name in always-loaded rule context; planner-agent template
demonstrated the inline file content its own line 110 forbids
The sweeps land as tests/commands/command-shell-state-shape.test.mjs, verified
red before the fix and proven able to fail by reintroducing the defect. Two
existing tests asserted the old bash-block mechanism rather than the intent and
were updated. Suite 1449/0; frozen v5.0.0 snapshots and all scanner code
untouched.
Not fixed, deliberately: neither command scope-gates its actions to the audit
target. The generated plan included an edit to a real file under ~/.claude,
outside the throwaway target, because the skill/agent scanners are machine-wide.
That is a design change, not a side fix.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0195udHgCcFegzm7ecKku2Yc
Every command so far asked an addition question — what to add, what to move,
what it costs. Nothing asked what is no longer earning its always-loaded rent.
This adds that axis as a fourth lensCheck on the existing hybrid motor rather
than a new scanner or a 22nd command: the measured payoff (~18% of one file)
justifies a mode, not machinery.
It is the only lens that proposes REMOVING config, so it carries a guarantee
the others don't need: a load-bearing block is never a candidate. Precision is
asymmetric — a missed dead line costs a few tokens per turn, a deleted one
costs a wrong remote or a broken script — so the floor is decided in code
(lib/floor-exclusion.mjs) before the opus judge sees anything, never in prose.
Granularity is the leaf block, with two structural exceptions: a paragraph
ending in ':' merges with the list it introduces, and an ordered list is a
contract whose steps inherit floor from any sibling. Unordered lists
deliberately do not inherit — a load-bearing bullet and a disposable one
routinely share a list, and container-reasoning is the error the hand-built
ground truth exists to catch.
Verified against that ground truth (built before any classifier existed), with
the comparison machine-checked rather than read by eye: zero load-bearing
blocks proposed, 11/18 deletable groups surfaced, ~756 tok ~ 18% of a ~4300
token file — inside the pre-registered band. The first run found five floor
violations the synthesized fixture missed; each got a structural rule and a
fixture shape so it cannot regress.
Three real bugs the dogfood run exposed, all now covered:
- JS \b is ASCII-only, so /\bunngå\b/ never matches — every Norwegian keyword
ending in æ/ø/å was silently dead.
- A bare word/word is not a path; "pros/cons" vetoed the largest deletable
block until PATH_RE was tightened to rooted paths and globs.
- "Mid-sentence" must key on a preceding lowercase letter; the loose version
read **bold labels:** and quoted openers as entities, costing 4 of 11 groups.
BP-SUB-001 is grounded entirely in the Anthropic steering blog already cited by
BP-MECH-001..004 and asserts nothing from the talk that motivated the feature —
no "80%", no ablation figure.
Suite 1365 -> 1382/0. Frozen v5.0.0 snapshots untouched; plain optimize output
byte-identical on identical input (--subtract adds keys only when passed).
knowledge-refresh-cli's reference date moved to 2026-08-01: its premise that
every seed entry was verified 2026-06-20 expired when BP-SUB-001 got a genuine
verification date, and backdating the entry to fit the test would have been a
lie about when its source was checked.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RW2haJXbxZpKivKHseSXNh
implement.md spawns implementer agents in parallel batches, all appending to
the same implementation-log.md. Dogfooding showed agents satisfying 'Append
result to:' with a full-file Write — the last writer clobbered 4 of 6 entries.
Pin the mechanism in both contracts: append with Bash >> heredoc, never the
Write/Edit tool on the shared log. Shape tests pin the instruction in both
files (empirically verified: agents given the >> instruction appended safely).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The Claude Code subagent harness instructs spawned agents NOT to write
report/summary/findings/analysis .md files — the parent reads the final
text message. Verified live: analyzer-agent skipped Write entirely and
returned the report inline, so analysis-report.md never landed on disk
and the plan/interview/status phases would find nothing to read.
New contract (orchestrator-writes pattern): analyzer-agent returns the
complete report as its final message; the analyze command saves it
verbatim to the session directory before presenting the summary.
Same class exists in plan/feature-gap/optimize/scanner agent pairs —
deliberately left for their own dogfood chunks (plan is judged as-is
first per the pipeline sequence).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CTontYwY5JGS4nL2AuiASy
The recall+precision halves of the CA-OPT hybrid motor for the three
mechanism-fit cases the deterministic OPT scanner (2a) deliberately skips:
lifecycle→hook (BP-MECH-001), unscoped path-specific→rule (BP-MECH-002),
absolute "never"→permission (BP-MECH-004).
New:
- scanners/lib/lens-prefilter.mjs — cheap, recall-oriented line scan of the
CLAUDE.md body; detector names mirror the register lensCheck fields; skips
fenced code, gates the path class on an instruction verb. Pure + 13 tests.
- scanners/optimize-lens-cli.mjs — discovery + OPT scanner + pre-filter; attaches
only the CONFIRMED register entry to each candidate (unverifiable → dropped,
Verifiseringsplikt); emits {deterministic, candidates, register, counts}.
- agents/optimization-lens-agent.md — opus precision gate (7th agent, orange):
reads the real CLAUDE.md, drops low-confidence candidates, keeps only genuine
opportunities, cites register id + source.
- commands/optimize.md — /config-audit optimize orchestrates pre-filter→agent→report.
Agent-driven → deliberately NOT byte-stable (own command, outside the snapshot
suite). No new orchestrated scanner → scanner count stays 15. Counts: agents
6→7, commands 18→19, suite 1055→1068. Self-audit A/A unchanged, readmeCheck
passed (clean HOME). Plan: docs/v5.7-optimization-lens-plan.md.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Wave 5 Step 16 — final wave step. Threads humanizer-aware rendering
rules through the three agent prompts that produce user-facing output,
and adds a shape test that locks the structure.
- agents/analyzer-agent.md: documents the humanizer envelope shape
(userImpactCategory, userActionLanguage, relevanceContext) in the
Input section; new "Humanizer-aware rendering rules" subsection
instructs the agent to: render humanized title/description/
recommendation verbatim, group findings by userImpactCategory, lead
each line with userActionLanguage, surface relevanceContext when
not affects-everyone, and skip jargon-translation subroutines.
--raw fallback documented (v5.0.0 verbatim severity prefiks).
- agents/planner-agent.md: documents the same vocabulary; instructs
the planner to consume humanized fields from the analysis report,
preserve titles verbatim, and order actions by both dependencies
AND userActionLanguage urgency. Translation duties explicitly
removed from the plan.
- agents/feature-gap-agent.md: replaces the inline t1/t2/t3/t4
tier-to-prose section ladder with userActionLanguage-driven
groupings ("Fix soon" → High Impact, "Fix when convenient" →
Worth Considering, "Optional cleanup"/"FYI" → Explore When Ready);
instructs skipping findings whose relevanceContext is
test-fixture-no-impact; --raw fallback documented.
tests/agents/agent-prompt-shape.test.mjs (new, +6 tests, 786 → 792):
- structural: humanized field reference + frontmatter preserved
- per-agent anchors: analyzer groups by userImpactCategory; planner
orders by userActionLanguage; feature-gap references
test-fixture-no-impact
- global: no "explain what {jargon} means" / "translate jargon" /
"jargon-translation duty" prose anywhere
Self-audit: Grade A unchanged (config 97/100, plugin 100/100).
feature-gap-agent and /posture command both reference quality area count.
Update both to reflect Token Efficiency as the 8th area.
Tests: 543 passing.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add /ultraresearch-local for structured research combining local codebase
analysis with external knowledge via parallel agent swarms. Produces research
briefs with triangulation, confidence ratings, and source quality assessment.
New command: /ultraresearch-local with modes --quick, --local, --external, --fg.
New agents: research-orchestrator (opus), docs-researcher, community-researcher,
security-researcher, contrarian-researcher, gemini-bridge (all sonnet).
New template: research-brief-template.md.
Integration: --research flag in /ultraplan-local accepts pre-built research
briefs (up to 3), enriches the interview and exploration phases. Planning
orchestrator cross-references brief findings during synthesis.
Design principle: Context Engineering — right information to right agent at
right time. Research briefs are structured artifacts in the pipeline:
ultraresearch → brief → ultraplan --research → plan → ultraexecute.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
