Skip to content

skill-authoring §6: name the deployment runtime before the review concludes#52

Merged
F-e-u-e-r merged 5 commits into
F-e-u-e-r:mainfrom
firaen22:mine-deployment-runtime-lens
Jul 22, 2026
Merged

skill-authoring §6: name the deployment runtime before the review concludes#52
F-e-u-e-r merged 5 commits into
F-e-u-e-r:mainfrom
firaen22:mine-deployment-runtime-lens

Conversation

@firaen22

@firaen22 firaen22 commented Jul 21, 2026 •
edited by F-e-u-e-r
Loading

Copy link
Copy Markdown
Contributor

What

Adds one rule to skill-authoring §6: name the deployment runtime before the review concludes — ask where the skill/plugin will actually run as a review input, then sweep for machine-bound assumptions (absolute paths, OS-specific launchers, host identity, clock/timezone, hardcoded MCP tool-name prefixes). Done when the review names the target runtime(s) and each machine-bound assumption is portable or labeled.

Why

§6's lenses (factual / doctrine / usability) all check the file against the author's world; none asks where it will run. A private incident showed the cost: a skill authored, reviewed, and finalized for the author's local macOS was rebuilt wholesale the same week when its real target — a sandboxed Linux VM — surfaced after sign-off; host identity, launcher mechanism, and machine-local MCP tool names all failed there. The question costs one sentence at review time and is a rework at deploy time.

Covenant status

Private incident cited as shape (contributor-verifiable, not linkable). Ships unprobed; Provenance records the probe shape (machine-bound skill + named foreign runtime, observe whether the sweep fires) as debt.

Checks

  • Single-file diff (grown across review rounds; see trail comment)
  • Unicode sweep of added lines: house-style §/—/✅/❌/→ only
  • Target runtime of this artifact: runtime-agnostic (pure doctrine text; the sweep found nothing machine-bound)
  • python3 .github/checks.py: all checks passed

🤖 Generated with Claude Code

firaen22 and others added 5 commits July 22, 2026 00:42
@firaen22
skill-authoring §6: name the deployment runtime before the review con…
4a6c54b 
…cludes

A skill reviewed and finalized against the author's local macOS was rebuilt
wholesale the same week when its real target — a sandboxed Linux VM —
surfaced after sign-off: host identity, launcher mechanism, and
machine-local MCP tool-name assumptions all failed there. §6's lenses check
the file against the author's world; none asks where it will run. This adds
that question as a review input plus the sweep list for machine-bound
assumptions (absolute paths, OS launchers, host identity, clock/timezone,
hard-coded MCP tool prefixes). Ships unprobed per the README covenant;
Provenance records the probe shape as debt.
@F-e-u-e-r @claude
skill-authoring: rework the deployment-runtime rule per r1 dual review
360c89d 
PR F-e-u-e-r#52 round-1 gate (grok-4.5 high: FIX 1,2; gpt-5.6-sol max:
FIX 1-5). Reproduced and addressed:
- unknown/unaskable runtime -> mark user-must-provide, conclude
  without a portability verdict, never default to the authoring host
  (grok F1)
- MCP clause made executable and bounded to instance-specific
  identifiers: verified portable name, or a runtime/instance label;
  no invented discovery mechanism (grok F2 / codex F5)
- Done now requires compatibility with every NAMED target - a labeled
  incompatibility blocks completion or explicitly shrinks the
  supported scope (codex F1, F4)
- S2 citation restored to its qualified form: hard-coded
  machine-absolute paths (codex F2)
- pos/neg pair added at the portable-vs-labeled boundary (codex F4 /
  grok n4); sweep categories glossed with concrete cues (grok n3);
  lead state-phrased (grok n5)
- codex F3 rejected-with-reason: single-incident rules with unprobed
  markers are this repo's documented covenant and settled precedent
  (silent-clobber, fan-in, interactive-runtime all shipped on one
  observation); S4's two-strike governs fix-log promotion in consuming
  projects, not this pack's covenant-governed rule adoption
- codex n6 (in-packet checks.py unverifiability): standing disposition,
  gate-runner re-runs it each round

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
@F-e-u-e-r @claude
skill-authoring: r2 rework of the deployment-runtime rule - always-sw…
d913732 
…eep, authoring-time ask, adoption block

PR F-e-u-e-r#52 round-2 gate (grok-4.5 high: FIX 1,2,3,4; gpt-5.6-sol max:
FIX 1,2,3,4,5,7). All reproduced and addressed:
- the sweep now runs on every review; naming is a conditional first
  step, not the trigger - an already-named target no longer skips the
  whole rule (grok F1 / codex F2)
- the runtime ASK moved to authoring start per the file's own
  gate-placement rule; this review verifies it landed (codex F1)
- no-answer path: user-must-provide recorded, sweep still runs,
  adoption blocked until the owner resolves or explicitly defers
  (grok F2 / codex F4)
- review record defined via S7's change-record shape; target runtime
  defined as execution environment + governing connector/tool instance
  (grok F3 / codex F7)
- machine-absolute paths get S2's VCS-root remedy - a label never
  keeps one (grok F4)
- scope shrink can only exclude an OPTIONAL target with the user's
  explicit say; labeled incompatibility with the actual deployment
  target blocks (codex F3)
- pos example made concrete and verifiable: named replacement, run on
  the target image, instance named (codex F5 / grok n5)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
@F-e-u-e-r @claude
skill-authoring: r3 refinements to the deployment-runtime rule
041b853 
PR F-e-u-e-r#52 round-3 gate (grok-4.5 high: FIX 1-5; gpt-5.6-sol max: FIX 1-8).
All reproduced and addressed:
- unknown-target semantics: requester supplies or records explicit
  risk-acceptance deferral; silence is neither; empty target list never
  satisfies Done vacuously (grok F1-analog / codex F1)
- named-target attribution: obtained from the requester, not
  author-inferred presence of a string (grok r3 F1)
- runtime-agnostic recorded outcome for doctrine-only files (grok F4)
- compatibility judged per target over assumptions REACHABLE on it;
  verified target-scoped dispatch branches count compatible (codex F2)
- S2 path remedy quoted verbatim incl. prefix verification, scoped to
  accidentally machine-local repo paths; target-defined absolute paths
  are ordinary machine-bound assumptions (codex F3, F6 / grok F2-analog)
- S7 record-selection clause quoted verbatim (codex F7)
- labels live IN the skill file beside the dependency, verified where
  reachable, marked unverified where not - a label records a
  limitation, never proves compatibility (codex F4, F8)
- review-cannot-retro-place-the-ask stated; S6 is the blocking
  backstop, not a time machine (codex F5); target-image run dropped
  from the pos example (grok F3-analog)
- optional-vs-required target pinned to the requester's explicit say
  (grok F5-analog)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
@F-e-u-e-r @claude
skill-authoring: r4 refinements to the deployment-runtime rule
ae73e9b 
PR F-e-u-e-r#52 round-4 gate (grok-4.5 high: FIX 1,2,3; gpt-5.6-sol max:
FIX 1-6). Adjudication:
- deferral terminal state unified: recorded risk acceptance by the
  deployment owner = alternate Done that still requires the sweep and
  in-file labels; no acceptance -> blocked; anti-vacuity clash removed
  (grok F1 / codex F1)
- target-scoped dispatch satisfies a named target only when that
  target keeps a working path for every claimed capability - costume
  conditionals excluded (grok F2)
- Done vocabulary unified on named-target-in-record (grok F3)
- sweep list declared a floor; executable-dependency dimensions added
  (arch, interpreter/deps, versions, fs, permissions, network);
  runtime-agnostic restricted to pure instruction text (codex F2)
- authoring-start requirement made answer-or-user-must-provide BEFORE
  first artifact-producing step; review = enforcement backstop
  (codex F4)
- plugin branch narrowed to a plugin's instruction files, matching the
  frontmatter router (codex F6); repo-manifests-first discovery order
  (codex n7)
- codex F3: this PR's own record gains its runtime line (body edit:
  runtime-agnostic, pure doctrine); codex n8: body Unicode claim
  corrected to include the arrow glyph
- codex F5 rejected-with-reason (standing): single-incident rules with
  unprobed markers are this repo's covenant and precedent; S4
  two-strike governs fix-log promotion in consuming projects

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
@F-e-u-e-r F-e-u-e-r mentioned this pull request Jul 22, 2026
Merged
F-e-u-e-r added a commit that referenced this pull request Jul 22, 2026
@F-e-u-e-r @claude
Fold combined-review r3: grok F1,F2 + n3-n6; codex F2,F3,F4 (scoped)
ed47089 
grok r3 F1: the labels rule's unverified-route block no longer bricks
wrappers with no attribution channel by design — channel present but
report missing/ambiguous/fallback-named blocks dispatch; no channel by
design yields a reachability-only pass with the limit recorded wherever
cited, never a verified-route claim.
grok r3 F2: pinned-string Done and pos example now carry the r2
route-attribution requirement (an unattributed probe never satisfies
the citation; the pos example shows the wrapper's route line).
grok r3 n3: hosted-behavior carve-out parenthetical now names both
duties (date-stamp the recorded claim; decision-time re-probe); the
writing imperative binds explicitly to tool-interface negatives.
grok r3 n4: check-name trigger broadened from 'safe' to any cited
property of a change (safe, correct, covered).
grok r3 n5: port recovery clause moved inside the address-in-use
branch; other bind errors stay terminal for that path.
grok r3 n6: ledger reference now cites the field by its actual title
('Recurring dispatches carry ledgers').
codex r3 F2 (scoped): attribution channel defined as reporting what
ANSWERED — a banner echoing the requested slug is configuration, not
attribution.
codex r3 F3 (scoped): the re-verify probe example now exercises the
claimed-absent capability; runtime-scoped verification remains S6's
deployment-runtime gate (no duplicate home).
codex r3 F4: ledger write-back gains the regressed-prior-fix
transition — re-flagging with evidence spawns a new OPEN finding, the
historical entry staying put with a pointer.

Rejected with counter-evidence (recorded in the PR trail): codex r3 F1
(mutate-and-restore on an unenforced copy is the exact case the
reference already rules — 'an unenforced reader can mutate and restore
without a trace; no endpoint comparison proves the read stayed clean'
-> provisional, never a clean gate pass); codex r3 F5 (re-raise of
PR #52's two-strike standing rejection — S1 placement already gates
authoring start, S6 is the named enforcement backstop, adjudicated in
that PR's own review; the frontmatter loader is pre-existing structure
outside this diff).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
@F-e-u-e-r
F-e-u-e-r merged commit f9cbe24 into F-e-u-e-r:main Jul 22, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@firaen22 @F-e-u-e-r