Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Original file line number Diff line number Diff line change
@@ -0,0 +1,8 @@
---
bump: patch
---

### Changed

- The corrector skills' instructions drop capital-letter emphasis and state each rule once, with no "when in doubt" default that widened subagent dispatch.
- Stale claims are corrected: the README lists the `setup` skill and its checkpoint behavior, `do-your-research` names the `strict` verification level as a reason for the full tier, and observed product behavior points at its live source.
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
---
bump: patch
---

### Changed

- The explore, research and blindspot instructions drop capital-letter emphasis, settle rules that disagreed between files, and correct stale section names and minimums.
Original file line number Diff line number Diff line change
@@ -0,0 +1,8 @@
---
bump: patch
---

### Changed

- The clean and setup instructions, and the engine's handoff-verify and alias-probe messages, drop capital-letter emphasis; every safety gate and approval rule is unchanged.
- The safety model and scan-flag references state each rule once and point at the Python and PowerShell docs instead of restating them.
Original file line number Diff line number Diff line change
@@ -0,0 +1,9 @@
---
bump: patch
---

### Changed

- The audit, compress, extract-ssot, rename-references and write-for-agents instructions drop capital-letter emphasis and settle points where two files disagreed, such as when the semantic-diff pass runs and whether extract-ssot commits its own waves.
- extract-ssot's identify survey and verify Gate 1 both leave generated blocks and the candidate SSOT's own file out of reproduction counts.
- Stale line-number citations, undated measured figures and restated tool behavior are replaced with current values or pointers to the live source.
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
---
bump: patch
---

### Changed

- The setup skill says plainly which action interviews for a missing key, and states resolver behavior in the present tense.
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
---
bump: patch
---

### Changed

- The grounding and update instructions drop capital-letter emphasis.
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
---
bump: patch
---

### Changed

- The teach, quiz-me and setup instructions drop capital-letter emphasis.
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
---
bump: patch
---

### Changed

- The README and setup skill name both configuration options the plugin declares.
Original file line number Diff line number Diff line change
@@ -0,0 +1,8 @@
---
bump: patch
---

### Changed

- The design and methodology instructions drop capital-letter emphasis.
- The price-before-running grader's fail clause now matches its pass clause and the case's expected outcome: 24 runs and a 4.2 to 5 USD band.
Original file line number Diff line number Diff line change
@@ -0,0 +1,8 @@
---
bump: patch
---

### Changed

- The simulation references settle counts that disagreed between files (personas, events per persona, focal moments, legend size, sticky spacing) and use a random temp directory for the extracted book.
- The methodology skill uses the notation table's colors throughout, drops a fixed search quota, and names the run-state file the iteration workflow actually writes.
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
---
bump: patch
---

### Changed

- The produce skill runs the stills script by its plugin-root path, like every other script it calls.
Original file line number Diff line number Diff line change
@@ -0,0 +1,8 @@
---
bump: patch
---

### Changed

- The README gives the current subcommand count, and the command reference sends the reader to the CLI's `--help` for `parse` file limits and refund amounts instead of restating them.
- `--pretty` is no longer listed as a global flag (`search` has none), scheduling advice points at the Claude Code scheduling comparison, and history notes about the earlier MCP setup are gone.
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
---
bump: patch
---

### Changed

- The reach skill's instructions drop capital-letter emphasis.
Original file line number Diff line number Diff line change
@@ -0,0 +1,8 @@
---
bump: patch
---

### Changed

- The DLSS 5 references date the fork-activity claim and agree with the skill on where `refetch` records its results.
- The setup skill's instructions drop capital-letter emphasis.
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
---
bump: patch
---

### Changed

- The audit skill describes all three conventions layers, the personal override included, and the security-posture recipe states answering every applicable item as a requirement rather than a hedge.
Original file line number Diff line number Diff line change
@@ -0,0 +1,7 @@
---
bump: patch
---

### Changed

- The README no longer names `spelling` as a formatter that needs repository configuration, and the setup skill drops capital-letter emphasis.
Original file line number Diff line number Diff line change
@@ -0,0 +1,8 @@
---
bump: patch
---

### Changed

- The plugin description counts 18 guards, the two advisory guards' option descriptions say only that they are off by default and how to turn them on, and the README states where the reference verifiers run and how hook budgets are set.
- The root-delete guard's block messages name the real cause of an unfollowable `cd` and say where a literal path goes for robocopy and .NET delete calls; the setup skill says which actions ask questions.
2 changes: 1 addition & 1 deletion docs/catalog.md
Original file line number Diff line number Diff line change
Expand Up @@ -107,7 +107,7 @@ plugin manifests and kept in sync by CI. Never hand-edit it; the category vocabu

## Security

- [`guardrails`](../plugins/guardrails): 17 toggleable guards. Blocks: secret writes, machine paths, git hook bypass, force-push, reset --hard, open-review merges, cross-domain paths, PR-fix requests, drive-root /tmp writes, exported MSYS_NO_PATHCONV, cd && git, foreground waits, root/out-of-tree recursive deletes, shell-expanded commit -m, off-convention commit/PR titles; off by default, credential reads. Warns: Workflow fan-out, bare gh pr create. Git hooks check commit subjects, secrets, paths, CLI flags, skill refs, removed paths.
- [`guardrails`](../plugins/guardrails): 18 toggleable guards. Blocks: secret writes, machine paths, git hook bypass, force-push, reset --hard, open-review merges, cross-domain paths, PR-fix requests, drive-root /tmp writes, exported MSYS_NO_PATHCONV, cd && git, foreground waits, root/out-of-tree recursive deletes, shell-expanded commit -m, off-convention commit/PR titles; off by default, credential reads. Warns: Workflow fan-out, bare gh pr create. Git hooks check commit subjects, secrets, paths, CLI flags, skill refs, removed paths.

## Workflow

Expand Down
10 changes: 6 additions & 4 deletions plugins/discipline/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -356,7 +356,7 @@ context you were missing, talks in ASD-STE100 Simplified Technical English
(short sentences, one meaning per word, technical terms exact), and uses the
project's ubiquitous language, read through
`/domain-driven-design:ubiquitous-language lookup` when installed, degrading
silently when no glossary exists. Re-pitch means re-ground, not compress: it stays out of the batch
silently when no glossary exists. Re-pitch means re-ground, not compress. It stays out of the batch
sweep, and a shorter-but-blunter reply is the failure it exists to avoid.

```shell
Expand Down Expand Up @@ -446,9 +446,11 @@ of truth:

## Configuration

No persistent state. Each skill reads the conversation and the consuming
project's own instruction layer. `follow-our-standards` may fetch a remote
standards source when the consumer declares one and no local checkout exists.
The plugin keeps no state of its own, with two exceptions: `recheck-against-upstream`'s mid-run
checkpoint to a durable topic-memory slice, when one exists, and the repository file
`/discipline:setup apply` writes. Each skill reads the conversation and the consuming project's own
instruction layer. `follow-our-standards` may fetch a remote standards source when the consumer
declares one and no local checkout exists.

The correctors themselves are zero-config. Three skills expose optional
`userConfig` scalars. The `sweep-all` runbook adds three that
Expand Down
5 changes: 4 additions & 1 deletion plugins/discipline/context/fan-out.md
Original file line number Diff line number Diff line change
Expand Up @@ -21,7 +21,10 @@ fresh-context subagent; this file owns how that dispatch runs.
- **Throttle in bounded waves.** A sustained wide fan-out trips server-side
burst overload (529s) and loses agents mid-run, so cap concurrency to a
modest wave (roughly a dozen or fewer at a time) and process the inventory
wave by wave rather than launching one agent per item at once.
wave by wave rather than launching one agent per item at once. For how a
subagent's API error surfaces, read
[API errors in subagents](https://code.claude.com/docs/en/sub-agents#api-errors-in-subagents)
live.
- **Retry the failed subset only.** If an agent errors or times out, retry that
item once; on a second failure mark it unverifiable, an honest skip, never a
false pass. Never blind-re-run the whole fan-out to recover a few stragglers.
Expand Down
10 changes: 5 additions & 5 deletions plugins/discipline/context/re-anchor-audit-correct.md
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
# The re-anchor / audit / correct-forward loop

Shared by every skill in this plugin. Each corrector re-anchors ONE
Shared by every skill in this plugin. Each corrector re-anchors one
standing discipline, applies it to the current conversation, and corrects
what has drifted. This file owns the common method; each skill's `SKILL.md`
adds only its own delta: which discipline, where that discipline's source
Expand Down Expand Up @@ -34,11 +34,11 @@ Run these in order. Skip a step only when its input is genuinely absent

1. **Re-anchor.** Re-read the discipline's source of truth (the skill names
it) with fresh attention and treat it as active for the rest of the
task. Loading the skill IS the re-anchor. The point is not to reprint
task. Loading the skill is the re-anchor. The point is not to reprint
the rules but to make them govern the next actions. State in one line
that the discipline now governs the work.
2. **Self-audit the work in flight.** Walk back over the conversation and
name CONCRETE, located findings, not a generic mea culpa. Each finding
name concrete, located findings, not a generic mea culpa. Each finding
points at a specific turn, claim, or artifact and at the specific part
of the discipline it breaks. If there are none, say so plainly. Do not
invent findings to look diligent. For work that others depend on (code,
Expand All @@ -54,7 +54,7 @@ Run these in order. Skip a step only when its input is genuinely absent
batched fork) returns "clean (unconfirmed)"; the pass that dispatched it
runs the confirmation. A corrector whose step 2 already runs in a
fresh-context subagent (a declared delta) has met this.
3. **Correct forward now.** For each finding, do the missing work THIS
3. **Correct forward now.** For each finding, do the missing work this
turn rather than merely noting it: edit the file, fix the config,
re-derive the choice, in the working tree, now. Where your own
judgment is the suspected source of the drift, re-derive it in a
Expand Down Expand Up @@ -113,7 +113,7 @@ context, never from a baked-in path. Apply the resolution ladder:

1. **Declared → use it.** When the consuming project states the discipline
in its own instruction layer, whether its `CLAUDE.md`, its
`.claude/rules/`, or a team conventions doc it points to, re-anchor THAT
`.claude/rules/`, or a team conventions doc it points to, re-anchor that
text. It is already
in the model's context; the skill raises its salience and audits
against it.
Expand Down
2 changes: 1 addition & 1 deletion plugins/discipline/skills/be-ambitious/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -125,7 +125,7 @@ project's own instructions still govern.
report, run the outside check and say what it showed. A fresh-context
review stays a proposal under section 3.

## A declared species in this plugin: NOT a corrector
## A declared species in this plugin: not a corrector

The correctors re-anchor a standing discipline. This skill is a one-shot,
user-fired raise or deepen: only the user knows they want more than they
Expand Down
14 changes: 7 additions & 7 deletions plugins/discipline/skills/do-your-research/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,7 @@ Read it; this file adds only what is specific to research discipline.
Research and verification before assertion. Resolve its source of truth
per the method doc's ladder: if the consuming project states a
research/verification discipline in its own `CLAUDE.md` or `.claude/rules/`,
re-anchor THAT. Otherwise re-anchor this portable baseline:
re-anchor that. Otherwise re-anchor this portable baseline:

- **Assert nothing you cannot point to a source for.** A claim labeled
"known", "obvious", or "from memory" is unverified until a fetched
Expand Down Expand Up @@ -55,7 +55,7 @@ plugin is installed; failing both, the floor below. The floor is this skill's
own baseline, deliberately lighter than a research pass's numbers, because it
settles one claim mid-conversation:

- **Tier**. At least one source fetched THIS turn: the live environment, tool
- **Tier**. At least one source fetched this turn: the live environment, tool
output, or the upstream artifact itself. Recall, and a summary of a source
read in place of the source, are both below the floor. An upstream docs
page is read through the shared docs lookup
Expand All @@ -64,7 +64,7 @@ settles one claim mid-conversation:
`<scripts>` = `${CLAUDE_PLUGIN_ROOT}/scripts` and `<session>` =
`${CLAUDE_SESSION_ID}`), not a WebFetch summary of it.
- **Corroboration**, before a claim carries a decision, a second source from a
DIFFERENT upstream pool. One pool restated by three intermediaries is one
different upstream pool. One pool restated by three intermediaries is one
source; where no second pool exists, say that instead of counting the
restatements.
- **Recency**. For anything that ships releases, a check against the current
Expand All @@ -75,13 +75,13 @@ actually applied.

## Two directions: grounding, and checking what was already said

- **Preventive**. Grounding a claim BEFORE it is asserted. Skipping it ships a
- **Preventive**. Grounding a claim before it is asserted. Skipping it ships a
wrong claim.
- **Detective**. Checking claims ALREADY asserted, which is what "fact-check
- **Detective**. Checking claims already asserted, which is what "fact-check
that" asks for. Skipping it leaves a wrong claim standing while later work
builds on it.

Direction is not the tier boundary. DEPTH is. Both directions run in the
Direction is not the tier boundary. Depth is. Both directions run in the
inline audit and in the fan-out tier below.

## Audit. What to look for
Expand Down Expand Up @@ -133,7 +133,7 @@ step delta.

The inline audit above is the default. Take the fan-out tier instead when the
invocation argument is `tiered` or `full`, or when your own judgment is the
suspected source of bias across MANY load-bearing claims, or a request to
suspected source of bias across many load-bearing claims, or a request to
"fact-check everything" / "verify every claim" wants provable coverage of the
whole session. A single fact-check or a short session stays inline: there the
fan-out's subagent cost buys nothing.
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -2,7 +2,7 @@

Read this once [`../SKILL.md`](../SKILL.md) has chosen the fan-out tier and resolved its depth.
It runs in place of the inline audit and correct-forward steps: same research discipline, but it
enumerates a TYPED FULL INVENTORY of the session's claims and has fresh-context subagents verify
enumerates a typed full inventory of the session's claims and has fresh-context subagents verify
them against primary sources, which a self-check in the context that produced the claims cannot
do.

Expand All @@ -12,14 +12,14 @@ do.
check in this context, or an already-cited source), and fan fresh-context subagents out only
over the load-bearing items. Most sessions want this: it spends the subagent budget where drift
actually matters.
- **full**. Subagent-verify EVERY inventory item, trivial or not. Reach for it when the cost of a
- **full**. Subagent-verify every inventory item, trivial or not. Reach for it when the cost of a
single wrong "trivial" item is high enough to pay for exhaustive independent checking.

Report which depth ran and why (argument / configured / default).
Report which depth ran and why (strict verification level / argument / configured / default).

## The fan-out

1. **Enumerate a typed full inventory.** List EVERY claim the session rests on, as a checklist,
1. **Enumerate a typed full inventory.** List every claim the session rests on, as a checklist,
each tagged with its type so coverage is provable. Not just the obviously load-bearing ones:
- **assumptions**. Unstated premises the work took for granted;
- **asserted facts**. Statements presented as true;
Expand All @@ -34,20 +34,20 @@ Report which depth ran and why (argument / configured / default).
exactly one row per item. Do not spot-check one, and do not silently drop an item as
"obvious", an obvious item is a `verified` ledger row, not an omission.
2. **Fan out, throttled. Scope set by the resolved depth.** Dispatch fresh-context subagents
(blind to the reasoning that produced each item) to verify each against a PRIMARY source, not
(blind to the reasoning that produced each item) to verify each against a primary source, not
the same recall that produced it. Under **tiered**, the fan-out covers the load-bearing items
while the trivial / non-load-bearing ones are resolved inline (and still get a ledger row);
under **full**, every item goes through the fan-out. Dispatch per
[`fan-out.md`](../../../context/fan-out.md): blind fresh-context subagents, bounded waves,
failed-subset retry.
3. **Match the method to the item type.** An externally-verifiable item (an asserted fact, a
concrete specific) resolves against a source or the live environment. An INTERNAL item (an
concrete specific) resolves against a source or the live environment. An internal item (an
assumption, or a load-bearing premise with no external referent) has no citation to fetch. Its
honest verdict is a fresh-context re-derivation or a flag for the user to confirm, never a
manufactured source. The 100%-coverage rule is coverage of the checklist, not a demand that
every row name a URL.
4. **Merge and correct.** Fold the returns together, correct every falsified or unbacked item
THIS turn, and surface anything that stays unverifiable rather than smoothing over it.
this turn, and surface anything that stays unverifiable rather than smoothing over it.
5. **Report a per-item ledger. 100% of the inventory.** One row per checklist item (no silent
drops), keyed by claim, each carrying:
- **verdict**. Verified / corrected / unverifiable, plus withheld for a recommendation row
Expand All @@ -59,7 +59,7 @@ Report which depth ran and why (argument / configured / default).
discipline. Resolve its source of truth by the shared method's ladder, the project's
`CLAUDE.md` / `.claude/rules/` notion of official / authoritative / trusted, degrading to the
portable baseline when none is declared;
- **consensus count**. How many INDEPENDENT authoritative sources agreed; a lone source is
- **consensus count**. How many independent authoritative sources agreed; a lone source is
weaker than a consensus, so note when only one was found;
- **recency**. For a fact that can go stale (versions, pricing, APIs, defaults), the date or
version the source reflects.
Expand Down
Loading
Loading