condux v2.18.0

• feat(condux): add blueprint design-mockup skill (#85)
• fix(condux): OpenCode routing enforcement, close 4 of 5 contract gaps from #37's triage (#38, #39)

@jabworks/condux@0.12.0

Minor Changes

• #85 a0cf2c5 Thanks @vi-hieu! - Add the blueprint skill — dependency-free design-time visuals: grayscale
  HTML wireframes for UI screens and flows, inline-SVG diagrams for data models,
  flows, architecture, and state machines. Loaded by discovery at the propose
  step, citable from draft-plan task cards, and triggerable standalone
  ("mock this up", "visualize the data model"). Also updates the workflow
  routing payload and skill table to carry the new member.

@jabworks/condux@0.11.0

Minor Changes

• 5513b83 Thanks @vi-hieu! - OpenCode now enforces /workflow as the entry point, matching Claude Code and Codex.

  Both other hosts get this from a SessionStart hook that injects
  skills/workflow/hooks/routing.md. OpenCode has no equivalent hook, so until
  now routing there fell back to catalog inference alone — the ~80% path the
  hook exists to replace (docket #38).

  The plugin's config hook now pushes routing.md onto config.instructions
  instead. Verified empirically against a live OpenCode install (opencode debug config, v1.14.48): a config-hook mutation to instructions reaches
  the fully resolved config, the same mechanism already proven by this
  package's skills.paths registration — unlike tools, which is folded into
  permission before the hook runs and silently drops any hook-side mutation.

  This is a deliberate cost tradeoff, not a free win: SessionStart fires once
  per session start/clear/compact, but instructions is ambient — re-injected
  every turn, permanently, in the same channel as the user's own AGENTS.md
  (~390 tokens/turn for routing.md). Chosen anyway for routing-enforcement
  parity with the other two hosts.

  condux-doctor's OpenCode probe now also checks that the installed package
  ships skills/workflow/hooks/routing.md, so a broken install is reported as
  broken instead of silently missing the enforcement.

Patch Changes

• 5513b83 Thanks @vi-hieu! - Four skill-contract gaps surfaced by the #37 sibling-miss triage are fixed (docket #39).

  root-cause-analysis now claims declarative bug reports, not just
  questions — "checkout crashes on empty cart" is a bug report even without a
  question mark, the same passive-voice-to-user-phrasing move that fixed
  test-first-development in 2.17.2.

  subagent-execution's when_to_use now names model selection for a
  dispatched agent explicitly — the skill already owned this in its body
  (references/spawn-rules.md), but the contract never said so, so "which
  model should the coder agent get" missed.

  remember (Concord) no longer attracts open-ended retrospective
  questions — "what mistakes did past sessions make", "what did the audit
  leave open", "has this happened before" in an unrelated project. Its
  contract now says explicitly: a session log, not a mistake ledger or audit
  index. A toolkit mistake in this project is still toolkit-failure-archaeology.

  The "sdd the plan" eval stimulus was genuinely ambiguous between
  spec-driven development and subagent-driven execution — reworded to "spawn
  the agents for this plan" rather than resolving the ambiguity in either
  skill's contract.

  git-operations' enumerated situation list (submodules, bisect) was left
  unchanged, by decision: the skill's own "Out of scope" section already
  excludes both by name pending a dedicated history-rewriting skill, so the
  miss is an intentional guard, not a gap.

toolkit-ops v1.7.5

• fix(evals): triage the sibling-miss cases — 19 accept alternates and one oracle flip (#37)

release v1.3.1

• fix(evals): triage the sibling-miss cases — 19 accept alternates and one oracle flip (#37)

condux v2.17.3

• fix(evals): triage the sibling-miss cases — 19 accept alternates and one oracle flip (#37)

condux v2.17.2

• fix(condux): let test-first-development claim the advisory questions and the fix-the-test requests it already owned (#37)

condux v2.17.1

• fix(condux): route subagent-deployment on the fan-out being asked for, and fix the eval oracle that punished the router (#32) (#81)

adapting-skills v1.4.1

• fix(evals): triage the sibling-miss cases — 19 accept alternates and one oracle flip (#37)

@jabworks/condux@0.10.2

0.10.2

Patch Changes

• e5403d1 Thanks @vi-hieu! - Eval-corpus triage for the condux member skills (docket #37, family A): seven
  cases gain accept alternates where the sibling the router actually chose is
  doctrinally correct — preflight/finalize/release on "ship it", preflight ↔
  live-verification on post-implementation "verify it", spec-browser on loading a
  spec, workflow on stimuli that are genuinely plain implementation requests.
  Each carries a note recording the rationale.

  No trigger contract changed. workflow chosen over a downstream condux skill
  deliberately stays a miss — those misses are the eval's power to detect drift
  toward the entry point, and they are the evidence for which downstream
  contracts are too weak to win.

• 6c9829b Thanks @vi-hieu! - test-first-development now claims the two things it already owned but never advertised.

  Advisory questions about the practice. "should I tdd ui components" missed
  0/3 in two independent variance bands — never once routed, the most stable miss
  in the corpus. The contract said "trigger when the user explicitly asks for
  tests-first", so a question about whether to read as a question, not a
  request, and the router declined. But deciding when tests-first applies is this
  skill's whole opt-in design. The contract now says so, mirroring workflow's
  "also the operating manual" clause.

  Requests to change a passing-by-editing test, in user phrasing. The rule was
  already there — "whenever an existing test spec is about to be edited to make it
  pass" — but written from the agent's side in passive voice, describing a state
  the agent is about to enter. The router only ever sees a user message, so
  matching it required a two-step inference it made about a third of the time.
  Named in user phrasing now: just fix the test, update the failing test to match
  the new behavior.

  No exclusion clause was added, deliberately. Across 582 cases, nothing wrongly
  routed to this skill in either band — zero false positives — so a "not for
  ordinary test work" clause would have bought nothing measurable (docket #37).