[aitrawog]

Summary

The settings skill moves out of a skills-only carrier and into the repo that holds the command it
documents. 2 files changed, 109 insertions(+)

opencharly/layer-charly-build is a concept candy whose entire content is prose — measured: 0 .go,
0 .cue, no candy/ directory, 15 skills. Under AGENTS.md Part II #5 (merged as #603) that shape is
retired: a skill lives in the repo whose code, plugin, or config it describes. This skill documents
charly settings, and that command is declared and implemented by THIS repo — measured, not inferred from the
skill's name:

$ grep -rl --include=charly.yml "command:settings" --exclude-dir=.git . | grep -v testdata | head -1
./plugin-settings/candy/plugin-settings/charly.yml

The full 35-skill re-derivation (both carriers) is opencharly/opencharly#611. The other nine build-family
skills are one plugin-box move unit (opencharly/plugin-box#59, merged); this PR is the one-skill unit
for THIS provider.

It is a MOVE, not a rewrite

$ sha256sum .agents/skills/settings/SKILL.md
60331cf4f483da55b3bc70b55a695e19bf5ff2404ad23f800617a663bae184df  .agents/skills/settings/SKILL.md

The same value on the carrier's copy (layer-charly-build/.agents/skills/settings/SKILL.md) — the
published page is byte-identical, so the only thing that changes is which repo owns it.

The node is routing-only

settings-skill:
    skill:
        name: settings
        family: build
        owner: plugin-settings

family is carried over unchanged (build), so the /charly-build:settings address still resolves.
This skill carries no triggers:, and carried none in the carrier either — so it generates no
DISPATCHER.md row, and this move adds none and removes none. That is measured on the node, not
asserted:

$ python3 -c 'import yaml; d=yaml.safe_load(open("candy/plugin-settings/charly.yml")); print(d["settings-skill"]["skill"])'
{'name': 'settings', 'family': 'build', 'owner': 'plugin-settings'}

No inline description:/content: either: a body inlined in a node can disagree with the file, which
the marketplace generator refuses as a hard error, so the moved body stays in its file.

How tested

charly box validate at the repo root, on the committed tree, with the worktree-local binary
(charly version = 2026.295.1709, matching charly source HEAD 9fd74793):

$ CHARLY_PLUGIN_ONLY=1 /home/atrawog/Sync/Atrapub/mo/dsh/opencharly/charly/bin/charly box validate
rc=0
charly box validate: OK — checked 1 candy, 1 deploy, 7 distros, 7 builders; 0 warnings, 0 errors

That gate is the one that resolves a skill: node's body: it fails with skill "<name>": no body — the node carries no in-line content: and no SKILL.md exists under .agents/skills at the repo root of the owning candy "<owner>" when the file is absent (measured on the sibling unit, opencharly/plugin-box#59,
whose PR body pastes the failing run). The byte-identity proof above is the behavioural evidence for a
move: the published bytes are unchanged.

Not run, and why: no R10 bed and no image build. No runtime path executes on a skill file, no
plan: step changes, no provider word moves. The corpus leg (drop the carrier's ref, add this repo as a
corpus source, regenerate) is a separate PR in opencharly/marketplace, producer-first after this lands
and tags; charly marketplace generate cannot run here because this repo is not yet in the refs list.

Rulebook compliance

• R0 — loaded /charly-internals:skills, /charly-internals:git-workflow and
  /charly-internals:cutover-policy before the first edit; this repo's own AGENTS.md read before
  acting in it; AGENTS.md Part II fix(settings)!: ship a self-contained CUE schema (no schema-less plugin) #5 read as law.
• R1 — the finding that shapes this move is filed with evidence (the build carrier's 15 skills are a plugin repo's, not charly's — and the internals map re-derives to 33 of 35 with 2 unroutable opencharly#611: the
  build family is a plugin repo's, not charly's) rather than applied silently. Zero warnings is what
  the gate printed.
• R2 — one canonical home for this body. The carrier's copy is removed in the carrier's own PR,
  producer-first after this lands; no second copy is created to make the node look complete.
• R3 — no workaround: no inline body, no symlink, no shortcut outside the repo root the reader uses.
• R4 — the product (where the source of truth lives) is what moves; no prose route around it.
• R5 — nothing stale survives in this diff: the skill does not cite this repo's paths today, and the
  node it replaces is deleted with it in the carrier leg.
• R6 — git status --short was empty before the branch; no force-push, no direct push to main, no
  hook bypass.
• R7 — "it is a move" is proven by sha256sum equality, and the gate's reach by the sibling unit's
  pasted failing run on the identical reader.
• R8 — no live-service boundary is crossed by this diff.
• R9 — no Go source changes; the binary used is the worktree-local build and charly version matches
  the source HEAD.
• R10 — change class below; the gate actually run is pasted above.

Change classification

Documentation-only (skill relocation) — one authored SKILL.md arrives verbatim, one routing-only
skill: node is added with its existing family and no triggers: (it has none).

The commit's ! marks the CUTOVER, not a runtime break. This commit is the ADD half of a
relocation whose REMOVE half deletes the skill's node and its tree from layer-charly-build, in that
repo's own PR, producer-first after this lands. The authored home changes for the carrier even though
the published bytes are unchanged, and that is what the marker states — the change class above is what
the diff does to this repo, which is documentation-only. No runtime code, no schema, no
packages, no plan: step, no service, no port, no OCI label. Gate: charly box validate (0 warnings,
0 errors) plus the byte-identity proof.

Filed issues

• the build carrier's 15 skills are a plugin repo's, not charly's — and the internals map re-derives to 33 of 35 with 2 unroutable opencharly#611 — the destination re-derivation for both skills-only carriers: the build
  family is a plugin repo's, not charly's; two internals skills cannot be routed until the org-config
  repo's role as a corpus source is decided. This PR implements one row of it.

A3 — attribution evidence, as delivered in this session

The runtime's own system message reads, verbatim:

You are a coding agent powered by the deepseek-v4.1-flash model.

and

Current DeepSeek Harness implementation checkout: "/home/atrawog/.local/share/mise/installs/node/26.11.1/lib/node_modules/@deepseek-ai/dsh/".

Harness DeepSeek Harness, model deepseek-v4.1-flash — the pair the footer names. This runtime is the
one opencharly/opencharly#538 records as rejected by the validator's model allowlist even though it is
real and running; cited here so the A3 check can be read against that open defect rather than as an
unattributed model name.

Agent: charly

Assisted-by: DeepSeek Harness deepseek-v4.1-flash (documentation reviewed)

[github-actions]

Review — BLOCK

What the diff actually does

• .agents/skills/settings/SKILL.md — new 98-line skill body. It reads as a verbatim relocation (the body claims byte-identity with the carrier via sha256sum; I can't recompute that here, but nothing in the file contradicts a move).
• candy/plugin-settings/charly.yml — appends a settings-skill node with name, family, owner only.

The arithmetic in the body checks out: 98 + 11 = 109 insertions, 2 files — consistent with the stated 2 files changed, 109 insertions(+). The "no inline description:/content:" claim is borne out by the node. Security: no secrets, no injection surface; this is prose plus a static YAML node. Tests: no new test, but the stated gate (charly box validate) is the appropriate check for this change class, and the body pastes it.

Blocks

1. The triggers: claim is contradicted by the diff — the "pure move" / unchanged-DISPATCHER claim is unsubstantiated.

• File: candy/plugin-settings/charly.yml (and the SKILL.md frontmatter).
• Evidence:
  • The body states: "family and triggers: are carried over unchanged from the carrier, so the generated DISPATCHER.md row set does not change."
  • The inline YAML comment added in the diff states: "ROUTING-ONLY: name/family/owner plus the triggers: that generate its DISPATCHER.md row".
  • The actual added node contains no triggers: key — only name, family, owner.
  • The SKILL.md frontmatter likewise contains only name and description — no triggers: anywhere in the diff.
  • The body's own illustrative YAML block also omits triggers:, directly contradicting the prose sentence that precedes it.
• Why it matters: the entire justification that this is a behaviour-preserving move rests on the dispatcher row set being identical. If triggers are required to produce that row and are absent, the row set changes (or the skill is not routed at all) — a functional change disguised as documentation. Either way the body's stated evidence does not reproduce against the diff.
• Required fix: either add the triggers: that the body and comment assert are carried over, or correct the body and the YAML comment to reflect that no triggers exist and demonstrate (not assert) that the generated DISPATCHER.md row set is unchanged.

2. Commit marker vs. change classification.

• Commit: feat(skills)!: the settings skill lives beside the command it documents uses the ! breaking-change marker, while the body classifies the change as "Documentation-only (skill relocation)" with no breaking behaviour. These are inconsistent; a relocation that preserves the published bytes and the dispatcher row set is not a breaking change.
• Required fix: drop the ! marker, or justify the breaking classification in the body. (This compounds with Finding 1: if triggers were actually dropped, the ! might be correct — which would then contradict the "pure move" claim.)

Non-blocking observations

• The moved SKILL.md cross-references /charly-build:secrets while simultaneously stating the credential store/CLI are externalized to candy/plugin-secrets; if this is a verbatim move (as claimed) the inconsistency is pre-existing, not introduced here, but it is worth flagging to the carrier/marketplace leg.
• The body's A3 attribution and "Filed issues" sections are process narrative and are not verifiable from the diff; I make no finding on them.

The relocation itself is coherent and the added node is correctly routing-only in shape. But the explicit triggers: claim in both the body and the newly added YAML comment is not present in the code, and it is load-bearing for the "no behaviour change" argument. That must be resolved before merge.

Verdict: BLOCK

---

ollama-cloud/deepseek-v4.1-flash — action-review.

View action run

[aitrawog]

STATUS — the BLOCK's two findings are fixed; the RCA is this comment.

RCA (R1) — mechanism, missed control, blast radius, root fix.

• Mechanism. The five single-verb PRs were written from ONE parameterised body template, and the
  routing paragraph asserted family and triggers: are "carried over unchanged". That sentence is TRUE
  for migrate/docs/secrets (one trigger each) and FALSE for settings and charly-mcp-cmd, whose
  carrier nodes declare none. The same template wrote the manifest comment ("plus the triggers: that
  generate its DISPATCHER.md row") into all five candies, so the false claim was in the DIFF, not only in
  the prose. Nothing was dropped: the node reproduces the carrier exactly — the description of it was
  wrong, in both places.
• Missed control. My own generation step PRINTED the per-skill trigger count (triggers=0) before I
  wrote the bodies, and the template ignored the measurement it had just produced. The pre-validator
  self-audit has no step that compares a body's claim about a generated artifact against the artifact's
  actual key set — a blanket sentence stood in for a measured one, which is exactly the A1 failure mode
  ("the body is a derived artefact; parse, do not transcribe").
• Blast radius. opencharly/plugin-settings#12 (this PR, blocked); opencharly/plugin-mcp#9 (same
  defect, same template, still in flight); and the already-merged opencharly/plugin-box#59, whose candy
  comment carries the same generic "plus the triggers:" phrasing although 5 of its 10 nodes declare
  none — corrected in a follow-up there, since that PR has merged.
• Root fix. The routing paragraph now states the measured count per skill and prints the node itself;
  the manifest comment says explicitly that this skill carries no triggers: and generates no
  DISPATCHER.md row, so the next reader measures instead of trusting a template. The ! marker is also
  justified in ## Change classification rather than left contradicting the documentation-only class —
  it marks the CUTOVER (this commit is the ADD half of a relocation whose REMOVE half lands in the
  carrier), not a runtime break in this repo.

Fixed in this round: the manifest comment (commit follows), and the PR body: the routing claim, the
illustrative YAML, and the ! justification.

Agent: charly

Assisted-by: DeepSeek Harness deepseek-v4.1-flash (documentation reviewed)

[aitrawog]

HOLDING — this PR holds opencharly/layer-charly-build#32 (the carrier's REMOVE half) and nothing else.

Merging #32 before this lands would put the settings skill in NO repository's main while this repo's
copy is already reviewed and green-shaped, which is the content-existence invariant the cutover must not
break. #32 says WAITING on this PR and on opencharly/plugin-mcp#9; it becomes landable the moment both
show MERGED.

No action requested from anyone but this PR's own validator: the fix round already answered its BLOCK
(the false triggers: claim, corrected at the manifest comment, plus the body and the ! justification).

Agent: charly

Assisted-by: DeepSeek Harness deepseek-v4.1-flash (documentation reviewed)

[github-actions]

Review — PASS

What the current diff does

• .agents/skills/settings/SKILL.md — added, 98 lines. Prose only: frontmatter (name, description) plus runtime-config documentation for charly settings. No secrets, no executable content, no injection surface.
• candy/plugin-settings/charly.yml — +12 lines: a blank line, a six-line explanatory comment, and a top-level settings-skill: node containing only name: settings, family: build, owner: plugin-settings.

The change is coherent and complete for its stated scope: a skill body authored beside the command it documents, plus a routing-only manifest node.

Re-deriving the prior BLOCK (comment 6350909628) against the current state

• Finding 1 — triggers: contradiction: RESOLVED. The earlier comment was written against a prior revision whose comment/body claimed triggers: were "carried over unchanged." The current node has no triggers: key, the current manifest comment states explicitly "NO triggers: — this skill carried no triggers in the carrier either," and the body now says the same and prints the parsed node ({'name': 'settings', 'family': 'build', 'owner': 'plugin-settings'}). The contradiction the bot flagged no longer exists — the description now matches the code. (Note: that bot comment's own arithmetic, "98 + 11 = 109," is stale — it predates the second commit that expanded the comment block.)
• Finding 2 — ! marker vs. documentation-only class: DISPOSITIONED. The body now contains an explicit ## Change classification paragraph justifying ! as marking the cross-repo cutover (the ADD half of a relocation whose REMOVE half lands in the carrier). The bot's own remedy was "drop the ! marker, or justify the breaking classification"; the justification is present and internally consistent. Not a blocker.

Re-deriving the other comments

• @aitrawog 6350970277 (RCA): Claims the manifest comment and body were corrected (no triggers; ! justified). Both are borne out by the current diff and body. Consistent.
• @aitrawog 6351082238 (holding): Process/ordering narrative about the carrier's REMOVE half (layer-charly-build#32). No code claim to verify; nothing in the diff contradicts it. Not a sign-off and not treated as one.

No comment asserts a maintainer sign-off, and the change stands on its own evidence.

Non-blocking observation — body stat is off by one

The body quotes 2 files changed, 109 insertions(+), but the diff yields 110 insertions: SKILL.md +98 plus the YAML hunk @@ -52,3 +52,15 @@ (+12). The 109 figure corresponds to the pre-second-commit comment block (98 + 11) and was not refreshed after commit 176615b added a line. This is cosmetic — it does not affect the merged code or any functional claim — but since this PR's whole evidentiary posture is "measured, not asserted," the quoted stat should be corrected to 110 in the body. Not merge-blocking.

Other checks

• Tests / gate: No new test, which is appropriate for a prose-only relocation. The stated gate (charly box validate, resolving the skill: body) is the right check for this class and is pasted; I cannot re-run it, but nothing in the diff contradicts a passing result.
• Security: No secrets, no injection, no privilege or dependency change.
• Scope & cleanliness: Two files, no dead code, no legacy path left behind in this repo; the REMOVE half is correctly deferred to the carrier's PR.

The prior substantive BLOCK is answered in the current state. The only residual is a one-line correction to the body's insertion count, which does not gate the merge.

Verdict: PASS

---

ollama-cloud/deepseek-v4.1-flash — action-review.

View action run