Re-measured the two committed chain cases under the stricter turn-boundary
scoring from 313a5809: 100% and 33% at RUNS=3, against 33% and 33% before.
That rise is NOT an effect of the change and the doc now says so. The turn
rule only ever rejects matches a flatter reading accepted, so it cannot
raise a score — the two runs simply straddle a 67-point swing on the same
prompt at RUNS=3, where one sample is worth 33 points. Reading it as an
improvement would be exactly the mistake the eval exists to prevent.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ABdTnqyBodw6mHVVkGehqe
Signed-off-by: cliffhall <cliff@futurescale.com>
One finding, and a sharp one: position in the stream is not causation.
`collectSkillInvocations` flattened every `tool_use` block across every
assistant event into one ordered array, and `chainHit` read position
alone. But the model can emit several `Skill` blocks in a SINGLE assistant
message, and it has not seen the first skill's body when it does — those
are parallel guesses, not a hand-off. A run that emitted `testing` and
`test-servers` together in one message scored as `A` led to `B`, and would
have gone on scoring after the pointer in `testing`'s body was deleted:
exactly the false pass this feature exists to rule out, arriving by a
second route (Copilot).
Each recorded invocation now carries the assistant event it came from —
`{payload, turn}` — and `chainHit` requires every link after the first to
land in a STRICTLY later turn. The turn count is threaded through
`collectSkillInvocations`'s `turnOffset`/`nextTurn` so a stream read in
chunks stays monotonic rather than restarting per read; without that, a
chunk boundary would reset the numbering and reintroduce the same bug on
long runs.
The scan stays greedy, which is still correct under the constraint: taking
the earliest occurrence of a link can only leave more room for the rest,
so no later starting point succeeds where the greedy one fails.
`sampleHit` is deliberately unaffected — a first-move case asks only
whether a skill fired at all, so it reads names and ignores turns.
Four new assertions in one case: a same-turn pair scores false; the same
two skills a turn apart score true; a same-turn pair does not poison a
later genuine hand-off; and only the chain's first link is unconstrained.
Plus a `collectSkillInvocations` case pinning the turn tagging and the
cross-chunk offset.
`docs/skill-authoring.md` now lists this as the second way a chained case
can false-pass, alongside a prompt carrying the target's own trigger.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ABdTnqyBodw6mHVVkGehqe
Signed-off-by: cliffhall <cliff@futurescale.com>
Two findings, both correct, and the first one means round 1's containment
fix was only half a fix.
**`--allowedTools` is not the availability filter.** It grants permission;
it does not restrict which tools exist. So a tool a user's or a plugin's
settings already permit stayed reachable across all 14 turns, and the
bound still rested on the deny list being complete — the exact property
round 1 set out to stop depending on. `--tools` is the real restriction
(it selects from the built-in set), and both flags now carry the same
four names: `--tools` to bound, `--allowedTools` so those four never need
a permission prompt no headless run can answer. Verified against the real
CLI — the full flag set runs to a clean `end_turn`. The test asserts both
flags, and both probe snippets in `docs/skill-authoring.md` carry
`--tools` with a ⚠️ saying which one does the bounding.
**The threshold guard let `NaN` and negatives through.**
`CHAIN_THRESHOLD=abc` is `NaN`, which fails every comparison and prints an
`above NaN%` summary; a negative bar passes every chain unconditionally.
Both turn an advertised env knob into a measurement that quietly means
nothing. `CHAIN_THRESHOLD` must now be in `[0, 1)` — half-open because the
chain bar is strict — and `THRESHOLD` in `[0, 1]`, inclusive because a
first-move floor of 1.0 is meetable. The message names the offending
value. A new case spawns the script for each bad value rather than
asserting on the predicate, since these guards live in `main`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ABdTnqyBodw6mHVVkGehqe
Signed-off-by: cliffhall <cliff@futurescale.com>
Five findings, all real.
**The chain prompts carried the target skill's own trigger.** `test-servers`
claims "a change needs a real server to exercise it", and both cases said
"against a real/live server" — so the model could load `testing`, then pick
`test-servers` from the ORIGINAL prompt, in that order, and score a hit that
would have survived deleting the pointer from `testing` entirely. Rewritten to
carry no server cue ("end to end"), and the measured rate fell from 100% / 67%
to 33% / 33%. That is the size of the artefact, and 33% is what the pointer
alone is worth. Both cases are now red against the 0.5 bar and stay that way:
`skills:eval` is not a gate, the number IS the finding, and lowering the bar to
turn the column green would discard the only signal this feature adds. Filed as
#2247 with the trap written down so a fix cannot re-measure the artefact.
**A deny list cannot bound a 14-turn run.** It only names tools known when it
was written, and this checkout configures an HTTP `mcp-docs` server in
`.mcp.json` while a contributor's own MCP servers and plugins add more. Now
`--allowedTools Read,Glob,Grep,Skill` (what the harness actually needs) plus
`--strict-mcp-config` with no `--mcp-config`, which drops every configured
server. The deny list stays alongside, since a deny is unconditional while an
allow list governs pre-approval. What remains outside all three is a
contributor's plugin tools; `--bare` would remove those and skills with them,
which would measure nothing — so that residual is stated rather than papered
over.
**The chain threshold compared inclusively.** "More often than not" is `> 0.5`,
and `>=` passes exactly half whenever RUNS is even — 2/4 would report a result
the criterion does not license. `passesThreshold` is now strict for chains and a
floor for first moves, and a strict bound of 1.0 is rejected up front rather
than failing every case while looking like a trigger problem.
**The doc's probe command still denied only four tools**, contradicting the
section's own promise that a probe and a scored case see the same policy.
Updated, with a note that it must move with the harness, plus a hand-off probe
form.
**The reporting had no automated coverage** — the one thing that could silently
merge the two measurements lived inside `main`. Extracted as `formatReport` and
tested: both headings, per-group thresholds, the two separate summaries, the
combined exit, and both single-kind selections.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ABdTnqyBodw6mHVVkGehqe
Signed-off-by: cliffhall <cliff@futurescale.com>
`skills:eval` ran every case with `--max-turns 1`, so it measured one
thing: whether a skill is the model's FIRST tool call. That is right for a
skill a user reaches directly, and it leaves a whole class unmeasured — a
skill reached from inside another skill's body. `testing` tells the model
that picking a fixture is `/test-servers` and that it has to load it, and
`test-servers` scores 5/5 while every one of those five cases asks for it
by name. A skill only ever reachable through a hand-off would score a
clean 100% with the hand-off silently never firing.
An eval case is now one of two shapes, and exactly one:
{ "prompt": "…", "expect": "test-servers" } first move, 1 turn
{ "prompt": "…", "chain": ["testing", "test-servers"] } hand-off, 14 turns
A chain must END with the skill whose file it lives in. The case exists to
measure whether THIS skill is reachable, so the file that goes red is the
one belonging to the skill that stopped being reached; anchoring on the
first link would file a `test-servers` measurement under `testing`, where
a `test-servers` description edit would never be seen. Links are checked
against the repo's model-invoked set, which is why both `verify:skills`
and `collectCases` now validate in a second pass — `test-servers` sorts
before `testing`, so a per-directory check would reject a live chain as
unknown purely because of where the alphabet put it.
Scoring is an ordered SUBSEQUENCE, not a prefix and not a contiguous run:
the model may load something before the chain starts and something
unrelated in between, and neither changes the claim that A led to B.
`collectSkillInvocations` therefore returns an ordered array rather than a
Set — a B, A, B run has to stay distinguishable from one that never
reached B from A.
The two rates are reported in separate columns and never summed. A
hand-off is a second-hop load over many turns; a first-move rate is the
model's opening move. `CHAIN_MAX_TURNS` (14) and `CHAIN_THRESHOLD` are
their own knobs for the same reason, and `CHAIN_THRESHOLD` defaults to
**0.5** rather than inheriting 0.8: see the measurement below.
Read-only containment no longer leans on `--max-turns 1`, which was doing
much of it by itself. The deny list gains the agentic and network tools —
`Task` in particular, whose subagent the flag does not reach.
Measured, `skills:eval -- test-servers`:
First move (1 turn) 7/7 at 100%, unchanged
Hand-off (14 turns) 67% and 33% at RUNS=3
So the hand-off is real and unreliable — which is the fact nothing could
observe before this change, and the reason 0.5 rather than 0.8 is the
default bar: at 0.8 both committed cases are red no matter how strongly
the first skill points at the second, and the column stops carrying
signal. A `["pr-flow", "test-servers"]` probe measured 0% and was dropped
rather than kept: `pr-flow` says nothing about fixtures, so the case had
no lever short of broadening a description onto another skill's ground.
Both findings are written up in `docs/skill-authoring.md`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ABdTnqyBodw6mHVVkGehqe
Signed-off-by: cliffhall <cliff@futurescale.com>
Restrict the truncation guard to PLAIN scalars. `yaml` attaches a legitimate
trailing comment to quoted and block scalars too, so keying on `.comment` alone
rejected valid frontmatter — `description: "Complete value" # note` was reported
as truncated, telling the author to quote a value that was already quoted. The
regex this replaced tested the same thing by reading the first character; moving
to the node dropped the check rather than encoding it. Verified all three
shapes (double-quoted, single-quoted, block) against the parser, and added the
regression cases.
Also correct the `--max-turns 1` bullet in docs/skill-authoring.md, which
claimed the skill must be the model's first tool CALL. `collectSkillInvocations`
scans every tool_use block in the assistant event without regard to order, so a
turn calling Read and Skill together scores a hit. What actually loses the
sample is needing a tool RESULT first — that decision lands in turn two, which
never runs. The practical advice is unchanged; the mechanism now matches the
collector.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SyBPs7caoMcyvnp9ArMBhV
Signed-off-by: cliffhall <cliff@futurescale.com>
Corrects the rationale for the five-positive floor, which was wrong. The
claim was that five prompts make a skill's rate flake-tolerant. The
evaluator scores each prompt on its own `passes / RUNS`, so the number of
prompts never enters any rate and cannot affect a threshold crossing —
`RUNS` is the only knob that does. The floor is worth keeping for
breadth: one prompt measures one phrasing, and a description that fires
on a single narrow shape would pass on its one lucky case.
The bad reasoning had reached five places — AGENTS.md, the authoring
checklist, the JSDoc, the check-site comment and the regression test —
plus the emitted message, which asserted it too. All corrected.
Moves `MIN_POSITIVE_CASES` above `validateEvalCases`'s JSDoc. It had been
inserted between the two, so the `@param`/`@returns` block documented the
constant and the function had none.
Replaces the two remaining cases answerable from `AGENTS.md`: `local-dev`'s
web-client-startup case (the rules file gives the exact command) and
`project-structure`'s clients/web-versus-core case (its tree answers it).
Both replacements probed 3/3 and measure 100%.
RUNS=5 skills:eval: 63/63, every local-dev and project-structure case at
100%. local:gate passes, 34/34.
Signed-off-by: cliffhall <cliff@futurescale.com>
Replaces the last `project-structure` case that `AGENTS.md` answers — a
helper used by the CLI and TUI belongs in `core/` per its own directory
tree. This was also the only case below 100% in the previous suite, so
the static reading and the score agreed. The replacement targets the
skill's first stated purpose, locating code, which `AGENTS.md` gives no
method for; it probed 3/3 and now measures 100%.
Extends the truncation guard's tests to `when_to_use`. The guard loops
over both fields but only `description` was pinned, so dropping the
second from that list would have gone unnoticed. The new case pairs a
quoted description with an unquoted `when_to_use` and asserts the
description is not implicated.
Scopes the negative-case requirement in the prose to model-invoked
skills; round 2 scoped the checklist but left this sentence saying
"every skill file".
Refreshes the stale `~2.8k of 4k` listing figure. The other instance is
past-tense about the three-to-nine experiment and stays, now said
explicitly so two numbers do not look like a disagreement.
Records the eval's real blind spot: all three cases review identified as
`AGENTS.md`-answerable were scoring 80-100% when flagged, so a high score
does not clear a case of overlap. The eval can show a case failing; it
cannot show a passing case is well-aimed.
RUNS=5 skills:eval: 63/63. local:gate passes, 34/34.
Signed-off-by: cliffhall <cliff@futurescale.com>
Enforces the five-positive standard this PR established rather than only
describing it. `validateEvalCases` checked `positives === 0`, so a skill
could regress to a single prompt and still pass the gate. At RUNS=5 one
case is 20 points of that skill's reading, so a one- or two-case skill
reports a rate a single flake can move across the threshold — well-formed
to the checker, meaningless as a measurement. Zero positives keeps its
own clearer message so the common mistake keeps its diagnosis.
Scopes the eval requirement to model-invoked skills in the checklist. A
name-only skill has no eval file because nothing can measure a trigger
only a human types, so an eval there is unrunnable rather than redundant.
Aligns checklist item 1 with the `#` hazard from round 1; it named only
the unquoted colon, which is the louder of the two failure modes.
Four test fixtures asserted on one-positive eval files and legitimately
went red; they now build positives from the exported MIN_POSITIVE_CASES.
The all-positive row keeps testing the missing-negative path, with a new
row for the below-floor case. 388/388 script tests pass.
No skill description or eval case changed, so trigger behaviour is
unaffected and skills:eval is unchanged from the previous round.
Signed-off-by: cliffhall <cliff@futurescale.com>
Fixes an unquoted `#` in `board-ops`'s description. YAML reads ` #28`
onward as a comment, so the parser — and Claude Code, which parses the
same way — only ever saw the description up to `Covers board`. Both
board numbers and the option-deletion hazard were absent from the
listing that decides whether the skill fires, and every check stayed
green because they look for an *empty* description and a truncated one
is still non-empty. Quoting restores all 271 characters, which is the
whole of the 3,144 vs 3,234 discrepancy the review flagged.
`parseSkill` now compares the raw scalar against what YAML kept and
errors on any loss, so the next one fails the gate rather than shipping.
Tests cover the truncated case and a legitimately quoted `#`.
Re-aims the `project-structure` Node-backend case, which `AGENTS.md`
answers via its tree and which had oscillated 80 -> 100% across runs —
the partial-overlap signature this PR documents. The replacement names
no internals and probed 4/4.
Makes the documented probe reproduce the harness: the same
`--disallowedTools` flags as `runPrompt` rather than `--allowedTools`,
`printf` instead of zsh's `print -r`, and `jq` over `tool_use` blocks
instead of a `grep` that also matched MCP server names in the init
event. The first-tool-call table is relabelled as an unrestricted probe,
since the eval cannot reach for Bash at all.
RUNS=5 skills:eval: 63/63, 62 at 100%; restoring those 90 characters
displaced nothing.
Signed-off-by: cliffhall <cliff@futurescale.com>
Adds a multi-name filter to `npm run skills:eval` so a set of related
skills can be iterated together, and makes an unknown name a hard error
rather than an empty run — a typo would otherwise report a green 0/0
that reads exactly like a clean pass of the skill that was meant.
Re-aims two `local-dev` cases and one `test-servers` case that partially
overlapped `AGENTS.md`. Partial overlap oscillates across runs rather
than sitting at zero, so it reads as eval noise instead of a defect in
the case; `docs/skill-authoring.md` now records that failure mode and
the observed scores.
Widens the `test-servers` description to the "Use when …" shape and
points the `testing` skill at it, since anything needing a real server
to run against is a procedure `testing` does not carry. That growth
moves the skill listing from 2,921 to 3,144 of 4,000 characters, so the
figure recorded in AGENTS.md is updated to match.
Signed-off-by: cliffhall <cliff@futurescale.com>
`skills:eval` was at 38/46 with several skills measuring 0-33%. The low
scores were mostly bad eval cases rather than bad descriptions, for two
reasons that were nowhere written down:
- The harness runs each case with `--max-turns 1`, so the skill has to be
the model's FIRST tool call. A prompt asking "how does X work here", or
naming a concrete file, sends the model to Read/Grep first — a better
answer to the question, scored as a miss. Every 0% case was one of these.
- `AGENTS.md` is resident on every turn, so a prompt it already answers is
answered without any skill and the case can never pass.
Replace the cases that measured the wrong thing, expand every model-invoked
skill from 3 to 5 positives (negatives kept) so one flake cannot cross the
threshold, and reshape `local-dev`'s description to the action-first form.
Add `docs/skill-authoring.md` with the case shapes that fire, the ones that
cannot, and the probe-then-measure loop.
RUNS=5 npm run skills:eval: 63/63 at or above threshold, 60 at 100%,
all 18 negatives clean.
Closes#2202
Signed-off-by: cliffhall <cliff@futurescale.com>
