Skip to content

v3.1.0

Latest
Latest

Choose a tag to compare

View all tags
@sprngr sprngr released this 30 Aug 04:52
· 1 commit to main since this release
v3.1.0
e767437

[v3.1.0] - 2026-08-27

Added

  • Session-start hook (opt-in --session-hook / -SessionHook): deterministically
    causes the rubber-duck agent to load the duck-policy skill at session start.
    • OpenCode: plugin installed to .opencode/plugins/session-start.js registers
      rubber-duck sessions and injects the startup directive into the system prompt
      on every model call (experimental.chat.system.transform).
    • Claude Code: scripts installed to .claude/hooks/ and a SessionStart hook
      merged into .claude/settings.local.json (idempotent). Fires only when the
      rubber-duck agent is active, detected via hook input agent_type.
    • Build emits hook artifacts to dist/opencode/hooks/ and dist/claude/hooks/.
    • Installer (bash + PowerShell parity): install/uninstall of hook artifacts,
      manifest pins, and per-target sessionHook tracking replayed on sync.
    • The rubber-duck agent body keeps its Enforcement Bootstrap mandate as a
      fallback when the hook is not installed.
    • Copilot support pending (deferred; see plan doc).
  • Reviewable-unit decomposition: multi-PR plans must decompose into reviewable units (independent merge, working state after each, explicit ordering + acceptance criteria). Methodology in src/shared/skill-snippets/reviewable-units.md; duck-design writes plans as PR sequences; duck-policy Checkpoint 1 gates on decomposition.
  • Validation fixture validation/fixtures/context-loading/src/cache.ts anchors deferred-debt marker tests to real code.
  • Plan decomposition verification spec: docs/architecture/08-plan-decomposition-verification.md defines the trigger predicate (explicit multi-PR OR agent-detected size/breadth), gate requirements, and verification acceptance criteria (positive, detection, negative, per-unit content).
  • duck-tidy extras skill (stub): audit-first cleanup for stale/outdated comments and non-CONTEXT docs. Installed with --extras / -Extras. Evidence rules: contradicts current code, describes removed behavior, worktree-only add/remove never merged. Carve-outs: TODO markers (duck-debt), ADR/design notes flag-only, CONTEXT.md/.duck-tape (duck-tape). Method bodies pending.
  • duck-adventure easter-egg skill (manual install): standalone rogue game — multi-turn dungeon crawls with maps, dice combat, random merchants, loot, and achievements tracked across sessions. Fun for its own sake; no handoff to productivity flows.

Changed

  • Duckling silent-worker contract: duckling subagent redesigned as single-turn silent worker across all three harnesses (Claude Code, Copilot, OpenCode). Interactive mutating-action-gate include replaced by explicit Silent Worker Contract:
    • one invocation is one turn; no mid-run user dialog
    • never self-approve mutating work
    • execute mode produces terminal approval package (preflight + per-file diffs + Approve this scope?) instead of mutating the workspace
    • mid-run ambiguity flattens to ## Unresolved questions block instead of interactive Q&A
    • one phase per invocation; parent orchestrates phase progression
    • tool-unavailable degradation emits explicit ## Tool unavailable note instead of silent skip
  • Duckling non-delegation list: duckling refuses skill_name of quack (routing skill), duck-tape (session-memory mutation), and duck-policy (session-scoped policy loader). Emits blocked_recursive_routing status; parent invokes target skill directly.
  • DUCKLING_CTX footer status vocabulary expanded: added blocked_awaiting_approval, blocked_skill_unavailable, blocked_recursive_routing, blocked_missing_inputs, phase_complete_await_parent, degraded_tool_unavailable. Prior <ok|blocked> binary replaced.
  • Rubber-duck Subagent Return Handling: new agent-body section defining shape-based and status-token-based recognition of subagent returns. Approval packages relayed verbatim as parent's Checkpoint 3 presentation; parent-always-executes rule (subagents propose; parent executes) preserves single-approval-gate invariant. Explicit handlers for phase-progression, blocked-input, and degraded returns.
  • Quack subagent-runbook: explicit disclaimer that quack has no return-side responsibility; parent (rubber-duck) owns return handling. Applies to any primary agent that dispatches to duckling.
  • Duckling general contract snippet (skill-snippets/duckling-general-contract.md): behavior rules 4-6 aligned with silent-worker posture. Interactive-dialog contracts (Socratic loops, batched interviews, multi-turn design dialogs) flatten questions to ## Unresolved questions. Skill-unavailable path emits terminal error instead of interactive question.
  • Duckling harness permission tightening: silent-worker contract now enforced at harness level in all three harnesses. Duckling tool maps drop mutation tools: Claude Read, Glob, Grep, Skill; Copilot read,search; OpenCode edit: deny, bash: deny. Duckling body gains explicit never-call-Edit/Write/Bash rule and parent-always-executes wording. duck-patch and duck-refactor gain "Subagent execution mode" sections documenting approval-package output under subagent invocation. Permission-tightening spike resolved inline.
  • duck-policy skill unchanged: portable policy layer stays agent-shape-agnostic; no duckling/quack-specific knowledge added.
  • OpenCode plugin uses system-level directive injection (system prompt) instead of
    user-message injection, so the model treats the directive as an instruction
    rather than a suggestion.
  • Sync replay fixes a latent off-by-one where extras were matched against the
    install-agents-md flag; extras and the new session-hook flag now map to the
    correct positional arguments.
  • Validation suite: 58 tests (V01-V58). V54 covers the Checkpoint 1 plan-decomposition gate; V55 positive trigger, V56 negative non-trigger, V57 per-unit acceptance content, V58 implicit detection trigger extend decomposition coverage via the rollout fixture. V25/V26/V38/V54 signals calibrated to observed vocabulary.
  • Validation runner now overlays built skills/ onto .agents/skills/ in test workspaces, so tests exercise current policy instead of the last installer-synced copy. Resolves false V58 failure caused by stale installed skills.
  • duck-policy Method adds gate-sequencing rule: approach-choice, clarify-first, Checkpoint 1 framing, and Checkpoint 2 fire as separate turns; clarify completes before framing. V51 prompt specifies the JWT failure mode so the Checkpoint 2 selection ask is deterministic.
  • docs/architecture/03-adaptive-socratic-policy.md Checkpoint 1 documents plan decomposition requirement.
  • Legacy managed-block migration no longer writes a .bak.<timestamp> recovery copy next to AGENTS.md/CLAUDE.md. The 3.x migration window is closed; the installer strips legacy blocks in place without backup (bash + PowerShell parity).
  • Rule wording convention applied across skills, agent bodies, and instruction docs: content-logic rules converted from absolute (always/never/must) to conditional if-then phrasing (~58 rules); structural/spec rules and safety carve-outs retain absolute or refusal form. Convention codified in duck-adapt (philosophy-core.md + adaptation-checklist.md).
  • duck-policy Style: gate ask strings are contract exceptions to terse style — Confirm or revise?, Select an option., Approve this scope?, Accept, revise, or rollback? emitted verbatim at their checkpoints even when otherwise terse. Mitigates intermittent gate compression (V50/V51). Version bump v3.0.1 -> v3.0.2.
  • duck-tape Resume: a marker from a previous session is a resumable handoff checkpoint, not a stop condition; stale markers now continue to checkpoint reload and position report. Version bump v2.1.2 -> v2.1.3.

Fixed

  • Validation runner: verdict matching evaluates the full multi-turn transcript instead of the final turn only; gate content from earlier turns (V50 Checkpoint 1 framing) no longer yields false negatives.
  • Validation runner: fixture workspaces no longer copy the rubber-duck AGENTS.md, removing workspace-identity confusion that intermittently made models ignore fixture code (V02/V07 class).
  • Validation tests calibrated to gate sequencing: V08/V25 gained follow_ups to advance past approach-choice/clarify-first; V45 steps through approach-choice -> clarify -> evidence -> Checkpoint 1 frame, signals reduced to problem+assumption (options coverage stays with V51).
  • Bash installer fails fast with a clear requires bash 4+ message (and macOS
    brew install bash guidance) instead of dying with an obscure
    declare: -A: invalid option on bash < 4 (macOS default /bin/bash 3.2).

Known regression

  • None active. Intermittent failures under flash-class models remain (LLM non-determinism, documented suite limitation): V51 (selection ask occasionally omitted), V02/V07 (workspace identity confusion — mitigated by the fixture-workspace AGENTS.md exclusion), V19 (bug-signal calibration on a bug-free diff), V33/V55 (judge variance), V45 (follow-up brittleness vs question phrasing). Mitigations: union matcher, gate-sequence test calibration, fixture-workspace AGENTS.md exclusion. Suite nominal 52/58 under opencode-go/deepseek-v4-flash; no Critical behavioral regression observed.
Assets 2
Loading