v6.1.9

What's Changed

• v6.1.9 by @github-actions[bot] in #159

Full Changelog: v6.1.8...v6.1.9

v6.1.8

v6.1.8 — 2026-06-17

Field-shape deduplication — shctx dups (#157). The third leg of the
mechanical shape-gate set (with dep-hygiene and check-impls-defs). Name-matching
dedup (index_symbols / dedup-check.sql / dedup_write_guard.sh / the conductor
DEDUP-GATE) catches a duplicate ONLY when the second definition reuses the first
one's name. It is blind to the dominant large-workspace rot: a second type for an
existing concept under a different name — the rename-to-evade-dedup shadow that
compiles green and slips every name-keyed gate (a 2026-06-17 audit of one workspace
found 22 such clusters). shctx dups closes that gap by clustering on field
shape, and gives subagents a way to recognize a pre-built struct they can reuse
without remembering its name — the match is surfaced, by shape, at authoring time.

New — shctx dups <scan|check|registry>

• scan — workspace census: parses every pub struct/pub enum, fingerprints
  each by its (field_name, normalized_type) set, clusters by similarity, and
  reports each cluster with members (file:line + consumer count), pairwise
  similarity, and a suggested canonical (lowest dep tier). --update persists
  the corpus to index_struct_shapes; --fail-on {medium|high|foundation-blocking}
  is a non-zero CLOSE/CI gate. The headline foundation-blocking severity flags an
  orphan canonical (zero consumers) sitting beside a live shadow.
• check <file> | --stdin --as <path> — authoring-time gate: matches a
  candidate's NEW defs against the corpus and reports any same-shape existing type
  ("…is 0.85-similar to pkg::X — reuse it?"). Exits 5 above the block
  threshold. Used by the PreToolUse hook and as a coder Phase-0 self-check.
• registry show|allow|unallow|pin|unpin|update — curated concept→canonical
  pins + a DO-NOT-MERGE allow-list for intentional distinct-role twins (a venue
  Fill vs a backtest SimFill). Tracked at <ns>/dups-registry.json.

Similarity, parsing, storage

• Metric: sim = name_weight·jaccard(field_names) + (1−name_weight)·jaccard(typed_pairs).
  The field-NAME blend catches a shadow that restated Uuid→String / DateTime→String
  / f64 field-for-field under a new name. Field-less (marker) shapes and shapes
  below dups_min_fields are excluded.
• Parser (skills/context/scripts/dups-core.py, stdlib python3) — a brace /
  generic / attribute-aware scanner over Rust source realizes the proposal's
  "Rust + syn" intent without a build step and is deterministic + unit-testable.
  Tree-sitter multi-language is a later extension; the shape model + similarity +
  clustering are language-agnostic. DB I/O routes through python's sqlite3
  module (no dependency on the sqlite3 binary).
• Schema: migrations/0015_struct_shapes.sql adds index_struct_shapes
  (the field-shape corpus, sibling of index_symbols/index_concepts).
• Refresh: shctx refresh --scope=shapes (folded into --all, hence
  sprint open) keeps the corpus current.

Enforcement + integration

• New PreToolUse(Write|Edit) hook hooks/scripts/dups_write_guard.sh — the
  shape-shaped sibling of dedup_write_guard.sh. @coder .rs writes only;
  config [dups].dups_hook = off | warn (default) | block. Fails open at every
  step (non-coder, non-rust, no python3, empty corpus → silent pass); it can only
  block on a real shape match.
• New doctrine doctrines/shape-dedup.md; zero-duplicate-tolerance.md gains
  Layer 4 + a cross-link.
• Config [dups] (docs/configuration.md + both example shepherd.tomls):
  dups_threshold, dups_block, dups_name_weight, dups_min_fields,
  dups_hook, dups_registry (keys are dups_-prefixed because cfg_get is
  section-agnostic). rust-service wires a shape-dedup close gate into
  [gates].extra.
• Tests: skills/context/tests/test_cmd_dups.sh (scan/check/registry/gate/persist)
  and hooks/tests/test_dups_write_guard.sh (block/warn/off + fast-paths), plus
  smoke cases.

Fix — the shctx absent false negative

A live v6.1.7 session reported shctx absent. Root cause: shctx is plugin-local
and never on $PATH — it is invoked by the absolute path
${CLAUDE_PLUGIN_ROOT}/skills/context/scripts/shctx. A command -v shctx /
which shctx probe returns absent by design, and when $CLAUDE_PLUGIN_ROOT
does not propagate into the agent's Bash env (some remote/web launches) even the
full-path invocation fails. Neither is evidence of absence.

• hooks/scripts/session_open.sh now surfaces, at SessionStart, the absolute
  shctx path resolved from the hook's own location (hooks/scripts → ../..),
  correct regardless of $CLAUDE_PLUGIN_ROOT, with the explicit note that
  command -v shctx returns absent by design. Config-gated
  [context].announce_shctx_path = on (default) | off.
• skills/context/SKILL.md documents the rule authoritatively (never PATH;
  command -v is the #1 false-negative; invoke by absolute path).
• Test: hooks/tests/test_shctx_locator.sh (surfaces the path with
  $CLAUDE_PLUGIN_ROOT unset; off-switch suppresses).

Fix — staged-handoff (v6.1.7) never actually worked

Verifying the v6.1.7 staged-handoff feature (/shepherd:spawn --staged +
seed-ready mailbox signal) surfaced a shipped defect: the mailbox.kind CHECK
constraint (from 0007) was a closed enum
(heartbeat_payload|escalation|ack|status|generic), so
shctx mailbox send --kind=seed-ready was rejected by the schema — the signal
could never be sent. Every future doctrine adding a routing tag would have hit the
same wall, silently.

• migrations/0016_mailbox_kind_relax.sql rebuilds mailbox with a permissive
  CHECK(kind <> '') (root-cause fix, not a one-value patch), preserving columns,
  data, the FK, both partial indexes, and the unread view.
• doctrines/staged-handoff.md jq consume snippet corrected to iterate the
  JSON array (.[] | select(...); recv emits an array).
• Test: skills/context/tests/test_staged_handoff.sh drives the full
  send → recv --unread-only --mark-read → ack seed-ready round-trip.

Both full suites green (hooks 44/44, shctx 44/44).

v6.1.7

What's Changed

• v6.1.7 by @github-actions[bot] in #153

Full Changelog: v6.1.6...v6.1.7

v6.1.6

v6.1.6 — 2026-06-15

Two-lane release. Lane A makes teammate-conductors actually leverage the
native Claude Code Workflow (Dynamic Workflows) tool instead of reflexively
ToolSearch-ing for it and mis-degrading. Lane B introduces /shepherd:ponytail
and the senior-engineering operating standard — the primitives that turn the
@auditor and @coder into senior devs, cemented into their profiles and adaptable
to the project and user.

The trigger for Lane A was a live failure: an /effort ultracode session, told to
"use the Workflow tool," ran ToolSearch select:Workflow, got nothing, and declared
the tool "confirmed absent" — the exact mistake teammate-conductors make. The fix
is not more prose (there was already plenty) but a recorded first-action
self-check that an agent cannot skip.

Lane A — teammates actually leverage Dynamic Workflows

• New doctrine doctrines/workflow-tool-self-check.md — the operational
  front-end for the native Workflow tool, consolidating the detect / never-
  ToolSearch / benefit guidance that was scattered across references/glossary.md §1,
  capability-discovery.md §V, workflow-compile-down.md, and
  primitive-axis-binding.md §IV (and still got skipped). ONE first action:
  is the token Workflow in your visible tool list? → record
  workflow_tool: present|absent → branch. Present → compile gate-free fan-out
  out-of-context (framed as the conductor's OWN benefit: clean context window + ≤16
  background agents, not a tax). Absent (web/remote, #146) → degrade to in-context
  Agent(...) — correct and expected. NEVER ToolSearch for it (the
  WORKFLOW-SELFCHECK-TOOLSEARCH anti-pattern), not even under /effort ultracode.
• Wired as a first action into agents/conductor.md (prohibition #22 rewritten +
  mode-comparison rows + a step 0 in the teammate compile sequence), agents/shepherd.md
  (root Step 0 + [ROOT-START] workflow_tool=…), commands/spawn.md (teammate boot
  prompt), and commands/start.md (lane walk step 0 + WAVE-COMPLETE.workflow_tool /
  .fanout fields). Corrected the stale "always present" phrasing to the
  environment-dependent visible-tool-list test.
• Enforcement seam: new @auditor completeness extension
  (agents/auditor.md §workflow-substrate discipline) — files PRIMITIVE-INVERSION
  when a lane hand-rolled in-context fan-out where the tool was present, and a LOW
  finding when the self-check was skipped or the tool was ToolSearched. In-context
  fan-out where the tool is absent is correct and never flagged.

Lane B — /shepherd:ponytail + the senior-engineering standard

• New doctrine doctrines/senior-engineering.md (the "ponytail" doctrine) — the
  senior-engineering operating standard for @auditor + @coder, eight primitives
  that build on the existing flock doctrines (cite, never duplicate): (I) comprehend
  intent before you touch (Chesterton's fence); (II) root-cause over symptom;
  (III) blast-radius- & cost-to-reverse-weighted severity; (IV) justify the tradeoff,
  not just the change; (V) conform to THIS project and THIS user via a precedence
  ladder (project doctrines > [CODE-STYLE] ledger > code-style skill > adaptation
  priors > the neighbors > defaults); (VI) cross-concern systemic-risk detection;
  (VII) bounded restraint (the most senior move is often the smaller diff);
  (VIII) preserved read-only/tier discipline. SENIOR-STANDARD-MISUSE anti-pattern.
• New command /shepherd:ponytail (commands/ponytail.md, also /ponytail) — an
  on-demand senior review → refine → verify pass on a target (diff, path, file, or
  PR) outside the sprint pipeline: Pattern 3 (Adversarial Verification) + Pattern 4
  (Generate-And-Filter), bounded as the AUDITOR-REFINE loop. Review-only by default;
  --apply runs the coder-refine + re-verify loop; --cement persists the conventions
  the pass observed into project memory / the style ledger. Fully adapted via styles,
  project doctrines, config, and adaptation priors.
• Cemented into the profiles: agents/auditor.md and agents/coder.md now cite
  senior-engineering.md in §"Doctrines this role honors" with the per-role primitive
  summary, and the FIRST-loaded reference skills (auditor.reference.md,
  coder.reference.md) carry the senior lens. New config section [ponytail]
  (senior_standard, default_mode, max_verify_iterations, apply_requires_approval,
  conformance_sources) in docs/configuration.md.

Wiring

• skills/shepherd/SKILL.md — /shepherd:ponytail + /ponytail triggers, command-table
  row, and §XI file-map rows for both new doctrines + the command.
• skills/shepherd/doctrines/README.md index — rows for workflow-tool-self-check.md
  and senior-engineering.md.
• CLAUDE.md command table + file-contracts; README.md command banner.

v6.1.5

v6.1.5 — 2026-06-15

Kickoff-hardening + config-auto-scaffold + observability release (#147), extended
with two new capabilities — #148 supervised self-heal and #146 capability
auto-discovery — plus a reliability follow-up that repairs the
operator-signaling inversion (the planter under-asked while the shepherds
over-asked), the "Workflow tool is always present" overclaim that made
web/remote sessions give up instead of degrading, and two latent namespace/DB
defects the new kickoff wiring exposed.

Authorized supervised self-heal — AUTONOMOUS-SENTINEL (#148)

• New loop template (references/loop-templates.md §AUTONOMOUS-SENTINEL) +
  binding doctrine (doctrines/autonomous-sentinel.md) for authorized
  supervised autonomy — the supervised-remediation superset of SOAK-LOOP.
  Stages: PROBE (seeded acceptance predicates, live) → CLASSIFY
  (HOLD/REGRESSED/NEW) → ACT (dispatch a ≤S @coder hotfix through the existing
  hotfix-dispatch ladder → gates-before-deploy → re-probe) → TERMINATE (K clean
  ticks / N-HF cap / hard-stop). Hard rails: gates-before-deploy, ≤S / ≤3
  concurrent / ≤N total HF caps, no destructive DB ops, auto-rollback on red,
  paper-only (never flip to live without authorization), operator-override-each-
  tick, full audit trail.
• New config key [close].autonomous_sentinel (default "off" — detection-only).
  It must be "on" AND the seed must declare close: autonomous-sentinel AND a
  complete sentinel_rails block must be present before a single remediation
  fires (three independent opt-in gates). New halt codes SENTINEL-RAILS-MISSING
  / -SCOPE-EXCEEDED / -HF-CAP / -ROLLBACK / -HARD-STOP / -LOOP-CAP.
• Reconciled the depth-3 "remediating inside a watch loop" anti-pattern in
  references/loop-templates.md §SOAK-LOOP and doctrines/outcome-enforcement.md §Seam 4: detection-only stays the DEFAULT and the anti-pattern for the
  UNAUTHORIZED case; the explicitly-authorized AUTONOMOUS-SENTINEL case is carved
  out.

Capability auto-discovery (#146)

• Shepherd now auto-detects the Claude Code plugins/skills available in the
  environment and adapts without operator wiring. A cheap, one-time-per-session
  SessionStart probe (hooks/scripts/capability_discovery.sh) enumerates
  installed plugins + skills and writes an EPHEMERAL capability roster
  (<ns>/cache/discovered-capabilities.json, gitignored) kept strictly distinct
  from the operator-curated toolkit.json — discovery never overwrites intent.
  The roster is merged at read time into the [TOOLKIT] surfaces (SessionStart
  roster + engineer/coder/planter brief injection via the new shctx toolkit discovered), labeled auto-discovered and bounded at 12.
• New doctrine doctrines/capability-discovery.md codifies the guarded-integration
  pattern ("if /remember is available → use at handoff/CLOSE-FINALIZE + resume,
  else shepherd-native"; same for superpowers, pr-review-toolkit), so behavior
  degrades cleanly when a plugin is absent — shepherd never hard-depends on a
  third-party plugin.
• The probe also records whether the native Workflow tool is present;
  web/remote sessions that omit it degrade to in-context Agent(...) fan-out
  instead of giving up (cross-referenced in references/glossary.md).
• New config key [discovery].auto_capabilities (on default | off), resolved
  via cfg_get (local → project → XDG-global precedence). Zero hot-path cost,
  fail-open.

Seed-optional kickoff (#8)

• /shepherd:start (Step 0) and /shepherd:spawn (Hard-stop #2 / Check 6) no
  longer hard-refuse on a missing seed for a single --scope sprint run: derive
  the objective from the repo/issue ledger, or ask ONE batched kickoff question,
  then run — per doctrines/operator-signaling.md §"Seed is recommended, not required". --parallel and multi-sprint --scope patch|minor|version walks
  still HARD-refuse (seeds are load-bearing there for collision detection + walk
  enumeration).

Config auto-scaffold (#15)

• New shctx config init scaffolds .claude/shepherd.toml from the bundled
  minimal template when absent (idempotent): derives [project].name (git
  remote → cwd basename) and [gates] (Cargo.toml→cargo, go.mod→go,
  pyproject/setup.py→pytest+ruff, package.json→npm), and realigns [paths] to
  the active shctx namespace. Adds shctx config get/show/path.
• Wired at kickoff: start/spawn root scaffold → [CONFIG] notice → PROCEED
  (action-biased); plant scaffold → ONE batched AskUserQuestion to refine
  [branching]+[gates] (replaces the #120 hard STOP).

Observability dashboard (#13)

• New shctx dash — a one-glance, read-only sprint snapshot composed from
  primitives the root already maintains (focus, graph state, live teammates,
  unread mailbox, open escalations, active loops, GitHub cache freshness). No
  new table/subsystem; bash-3.2-safe; degrades cleanly on missing DB/tmux.
  Monitoring recipe: /shepherd:loop <interval> shctx dash.

Four config toggles (#10)

• shctx config get <key> [default] is the uniform resolver (local→project→XDG)
  the toggles read through. Defaults reproduce pre-v6.1.5 behavior exactly:
  [autorun].on_grade_floor (abort), [autorun].inter_sprint_pause (brief),
  [spawn].max_parallel (4), [spawn].dashboard_cadence (3m). The
  previously-undocumented [autorun] section is now in docs/configuration.md.

Neutralized the bundled example (#9) + subagent-preference (#11)

• examples/axiom/ → examples/rust-service/; scrubbed all domain-specific
  references (finance/polymarket/geo-block) from the example and ~23 doctrine
  teaching snippets. geo-block-law.md rewritten as a generic
  regulated-upstream-API teaching example. Historical .artifacts/ docs are
  intentionally left intact.
• doctrines/agent-excellence.md Rule 6 (token-conservation / subagent
  preference) is now wired into every agent profile.

Reliability follow-up — operator-signaling inversion

doctrines/operator-signaling.md (v6.1.4) was correct, but its posture was
never reproduced into the agent profiles that actually become system prompts at
runtime. AskUserQuestion is granted correctly in every profile — the inversion
was a prose-propagation gap, not a tools-grant bug:

• agents/planter.md: added a standing "the planter asks freely" posture at
  the top of plant mode (previously the ONLY trigger was the rare no-config
  bootstrap branch, so the common case invented answers instead of asking).
• agents/conductor.md: SOLO = AskUserQuestion is a narrow escape valve only;
  TEAMMATE mode MUST NOT call it (the MODE-MISUSE halt code now names the
  tool).
• agents/shepherd.md: root action-bias note — the defined gates are the only
  operator stop points; no invented mid-run confirmation asks.

Reliability follow-up — the Workflow tool is NOT "always present"

• references/glossary.md listed the native Workflow tool alongside
  Agent/Bash/Edit as always-present and blamed any absence solely on "a
  build below the Dynamic Workflows floor." Claude-Code-on-the-web /
  remote-execution sessions omit it even on a supporting build, so a
  spawn/loop that reached for it gave up instead of degrading. Corrected to
  environment-dependent presence, with the visible-tool-list test as the
  only authority and degrade-to-Agent(...) as the documented path (ties into
  #146).

Reliability follow-up — two latent defects the kickoff wiring exposed

Both were invisible to the green suites (the harnesses set neither
CLAUDE_PLUGIN_ROOT nor a shepherd.db registry); #15's config-scaffold pulled
them onto the kickoff hot path:

• shctx_skill_root() returned bare $CLAUDE_PLUGIN_ROOT, but schema/ +
  references/ live at $CLAUDE_PLUGIN_ROOT/skills/context/. scaffold.sh's
  cp references/naming-conventions.md aborted under set -e, so shctx init/config init never created the DB and every downstream shctx command
  failed. Now prefers the dispatcher-exported SHCTX_SKILL_ROOT, else
  $CLAUDE_PLUGIN_ROOT/skills/context. Verified end-to-end (init exits 0,
  creates shepherd.db, copies CONVENTIONS.md).
• 9 hooks hardcoded $ns/root.db while v6.1.2+ shctx init creates
  shepherd.db, so [[ -f "$DB" ]] was always false → silent no-op, disabling
  the spawn-coordination guards (coordinate_drive_guard,
  worktree_teardown_guard, teammate_idle, …) on every modern project. New
  hook_db_path() in hooks/scripts/_lib.sh mirrors the skills-side
  shctx_db_path() (prefer shepherd.db, fall back to an existing root.db,
  default shepherd.db); all 9 assignments route through it.

Tests: hooks 38/38 (+1 for the #146 capability-discovery probe), context 42/42.

v6.1.4

What's Changed

• release: v6.1.4 by @github-actions[bot] in #144

Full Changelog: v6.1.3...v6.1.4

v6.1.3

v6.1.3 — 2026-06-12

The toolkit-hardening, bash-3.2-portability, and outcome-enforcement release. Fixes the v6.1.2 toolkit "Permission denied" that fired at session start, repairs three macOS bash-3.2 breakages (including a silently-broken hotfix guard and unbounded precompact-snapshot pileup), removes the retired autorun/parallel machinery for good, and adds a behavioral layer that makes the seeded outcome — not just green gates — the thing that closes a sprint.

Toolkit — the v6.1.2 feature, hardened

• The reported bug. hooks/scripts/toolkit_surface.sh and skills/context/scripts/cmd_toolkit.sh shipped in v6.1.2 with git mode 100644 (non-executable). The SessionStart hook is invoked by path, so it failed with Permission denied on every new session. Both are now 100755. A new regression guard — hooks/tests/test_exec_bits.sh — asserts every path-invoked hook/CLI script carries the executable bit (the smoke harness runs scripts via bash <file>, which is mode-agnostic, so this class slipped past every other test).
• CLI ↔ docs reconciliation. The documented add syntax was unusable (add <name> --desc=… --global while the CLI required --name=, --description=, --scope=global). The CLI now accepts the ergonomic aliases --global / --local (for --scope=…) and --desc (for --description); commands/toolkit.md and doctrines/toolkit.md are corrected to match.
• Doctrine accuracy. Dropped api from the canonical type enum (it is non-canonical → WARN); marked scope + capabilities required (the validate command enforces them); reworded the over-promising "pinned never drop" and "add is idempotent" claims to match the code (add refuses duplicates; pinned-first within the cap).
• Bounded injection + defensiveness. shctx toolkit md (the brief-injection surface) now caps at 12 entries, pinned-first, matching the SessionStart hook; interactive list stays uncapped. Empty-capabilities rendering is guarded. New hooks/tests/test_toolkit_surface.sh covers merge / local-wins / pinned-first / 12-cap / graceful-empty; test_toolkit.sh gains alias coverage.

bash 3.2 portability (macOS default /bin/bash)

Three hooks/CLI scripts used bash-4-only constructs that silently broke on the operator's platform:

• hotfix_vehicle_guard.sh (#135). Used ${SUBAGENT_TYPE,,} (bash-4 lowercase) → "bad substitution" on 3.2, so an H=1 hotfix spawned as a shepherd:conductor teammate slipped past the guard. Replaced with portable tr. The guard's primary case now actually enforces — test_hotfix_vehicle_guard.sh goes 9/10 → 10/10.
• precompact_snapshot.sh. The retention trim used mapfile (bash 4+); on 3.2 it died under set -u, so snapshots were written but never trimmed — the "so many precompact files" pileup. Replaced with a portable read loop; retention (default 5) now holds.
• cmd_issues.sh. declare -A (associative arrays) is a fatal "invalid option" on 3.2, breaking shctx issues outright. Reworked to safe printf -v / ${!var} indirection (no eval).

These two fixes turn the hooks smoke suite from 33/35 → 37/37 (the two newly-green tests plus the two new toolkit tests).

Precompact snapshots — relocated + meaningful

• Relocated from <workdir>/snapshots/ to <workdir>/memory/snapshots/, co-located with other ephemeral rehydration state. Writer (precompact_snapshot.sh) and reader (focus_rehydrate.sh) move together; a fail-open one-time migration sweeps any snapshots from the legacy location, and the reader falls back to it so an in-flight snapshot taken before the upgrade still rehydrates.
• Kept (they are meaningful): snapshots carry the focus digest, graph cursor, trace tail, mailbox, and lock state — the load-bearing compaction-resilience payload. The "noise" perception was the broken retention (above), now fixed.

Outcome enforcement — make the seeded outcome close the sprint

Shepherd ships code and green gates reliably but drifts off the outcome a seed promised. New doctrines/outcome-enforcement.md binds outcome verification to four existing seams, behavioral wiring only — no new schema or state table:

• Seam 1 — SEED: each seed §6 deliverable's acceptance is a runnable predicate (grep+count / structural assertion / LOC floor / log·metric·DB query / health probe), not prose (references/seed-template.md §6-bis).
• Seam 2 — PLAN-GATE: the @critic confirms every deliverable carries a runnable predicate; a plan that drops one or leaves prose-only fails with PLAN-MISSING-OUTCOME-VERIFICATION (agents/critic.md).
• Seam 3 — CLOSE (the enforcement point): the close auditor re-runs every seeded predicate against live HEAD/state before grading; a promised-true predicate that now returns false is an OUTCOME-REGRESSION (HIGH) that caps the completeness grade (agents/auditor.md, agents/conductor.md §3, references/grading-rubric.md). Reuses the same read-only re-run INTRO already runs on the prior sprint.
• Seam 4 — SOAK (optional, post-close): a new SOAK-LOOP template (references/loop-templates.md, a WATCH-LOOP specialization) re-verifies the predicates on a wall-clock interval (T+1d, T+7d) via native /loop + Monitor, surfacing OUTCOME-REGRESSION on post-delivery drift. Detection-only; never auto-remediates. Invoke: /shepherd:loop "soak outcomes for <sprint>" --agent worker --interval 1d --max 6.

Retired-command cleanup

Deleted the thin retired-redirect stubs (/shepherd:autorun + /shepherd:parallel, retired since v5.1.4): commands/{autorun,parallel}.md and skills/shepherd/{autorun,parallel,planter}.md. References in CLAUDE.md and SKILL.md are scrubbed; the live --auto / --parallel <N> flags, the [autorun].min_grade config key, and the shepherd-parallel-<slug> teammate-naming prefix are unaffected.

Doctrine + brief fixes (batched issues)

• #61 — agents/engineer.md gains a work-shape→vehicle tier-matching table so a conductor/teammate is not allocated for single-file or markdown work.
• #100 — agents/conductor.md boot adds a hard W0-GATE precondition: no body batch fires until INTRO ground-truth certification has passed (block-and-recheck, never proceed-and-hope).
• #120 — the planter (agents/planter.md, commands/plant.md) now has a fresh-project bootstrap: when .claude/shepherd.toml is absent it surfaces examples/minimal/shepherd.toml and stops for operator confirmation.
• #123 — reconciled the -dev.N seed-filename conflict: intermediate per-sprint seeds may carry -devN; the version-scale-roadmap restriction applies only to final shipped artifacts (doctrines/version-scale-roadmap.md, references/seed-template.md).

Closed as verified-shipped

Verified present + wired and closed: #107 (toolkit registry, v6.1.2), #121 (namespace resolution parity, v6.0.8), #134 (Focus Loop, v6.0.9), #87 (compile telemetry, v6.0.9), #103 (engineer model pin + ENGINEER-MODEL-FAIL, v6.0.3+), #99 (teammate git guard, v6.0.9), #119 (planter discovery wave, v6.0.7+), and #135 (hotfix vehicle guard — fixed and now enforcing, this release).

Test suite — repaired the shctx harness (23/40 → 40/40)

The skills/context/tests suite had 17 pre-existing failures (present on main) — not a sqlite issue (the migrations apply cleanly), but v6.1.2 renames that never reached the test harness:

• root.db → shepherd.db: shctx init now creates shepherd.db, but the harness (_setup.sh) and 22 tests still hardcoded root.db in their direct sqlite3 queries → "no such table" against a fresh empty file. Renamed throughout (leaving test_workdir.sh's intentional legacy-detection assertions).
• plans/+reports/ → docs/plans/+docs/reports/: test_init / test_lint referenced the pre-v6.1.2 top-level dirs the scaffold no longer creates.
• cmd_doctor.sh hard-coded the label root.db regardless of the actual file — now reports the real shctx_db_path basename (a real cosmetic bug).
• test_compile_telemetry asserted spaced JSON ("…": 1) against compact output ("…":1).

Both suites now pass clean: hooks 37/37, context 40/40.

---

v6.1.2

v6.1.2 — 2026-06-11

The self-improvement-substrate release: a persistent tool toolkit so a session never forgets a capability, a standardized + back-compatible workdir layout, per-flock-role loop templates, discovery waves on Dynamic Workflows, and a flock-profile polish pass.

Toolkit — persistent tool memory (operator request)

The flagship of this version. A mutable registry (toolkit.json) of commonly-used tools — MCP servers, skills, plugins, CLIs, ssh targets — so a Claude Code session never forgets a capability exists and the operator never has to re-explain it (e.g. ssh pzzld@laptop for a self-hosted dev surface, the context7 MCP). It is the tool-memory sibling of the adaptation loop's lesson-memory (doctrines/self-improvement.md).

• Two tiers, merged at read time. Project-local <namespace>/toolkit.json (tracked) ⊕ user-global $XDG_CONFIG_HOME/shepherd/toolkit.json; the scope field routes each entry, and local overrides global on name collision — so cross-project tools live once, globally.
• Entry schema. Required { name, scope (local|global), type (mcp|skill|plugin|cli), capabilities[], description } plus optional invocation, when, tags, pinned. JSON Schema at skills/context/references/toolkit.schema.json; the validator warns (never fails) on a non-canonical type so ssh/service targets are permitted.
• CLI. New skills/context/scripts/cmd_toolkit.sh (registered in shctx): toolkit list|add|rm|pin|unpin|show|md|init|validate. Lazily creates the file on first add; md emits compact markdown (graceful-empty — nothing on an empty registry, exactly like shctx adapt priors).
• Three surfaces keep it in front of the model. (1) A SessionStart hook hooks/scripts/toolkit_surface.sh injects a compact, ≤12-entry, pinned-first roster every session (fail-open; suppressed by [hooks].quiet_warnings); (2) the shctx toolkit CLI; (3) a [TOOLKIT] block injected into engineer/coder/planter briefs via cmd_inject.sh (variable-tail, cache-discipline-preserving).
• Doctrine + command + examples. skills/shepherd/doctrines/toolkit.md (bounded / graceful-empty / never-store-secrets), commands/toolkit.md (/shepherd:toolkit), examples/{axiom,minimal}/toolkit.json, and the five tool-using agents (engineer, coder, worker, discovery, planter) gained a one-line toolkit-awareness nudge. Test: skills/context/tests/test_toolkit.sh.

Standardized workdir layout — one consistent tree, totally back-compatible (operator request)

The per-project workdir now follows a standardized internal tree — docs/{plans,reports,diagrams,handoffs,specs,journal}/, logs/, archive/, cache/, scripts/, templates/, tmp/, types/, plus toolkit.json (tracked) and shepherd.db (gitignored). Adopted additively per the #121 "never mass-rename" invariant.

• root.db → shepherd.db, with auto-detection. shctx_db_path() prefers shepherd.db, falls back to legacy root.db, defaults to shepherd.db for new projects — mirroring the existing .shepherd/↔.artifacts/ resolution. Zero change for legacy trees.
• plans/ + reports/ now nest under docs/. [paths] defaults updated; scaffold.sh scaffolds the full tree with .gitkeep for tracked-but-empty dirs.
• Opt-in migration. shctx migrate --layout v2 git mvs plans/→docs/plans/, reports/→docs/reports/, renames root.db→shepherd.db, and creates the new dirs — idempotent, no-clobber.
• *.{group}.{ext} naming, formalized. references/naming-conventions.md documents the uniform <slug>.<group>.<ext> rule and adds log patterns {date}.log.md (human) + {ts}.log.jsonl (machine); cmd_lint.sh accepts both the legacy and new locations + log groups. .gitignore covers shepherd.db* under both namespaces and keeps toolkit.json tracked.

Per-role loop templates — bounded, role-shaped Loop-Until-Done (operator request)

/shepherd:loop (Pattern 6) gains a per-flock-role catalog so the loop primitive is reusable per agent. New skills/shepherd/references/loop-templates.md defines seven templates — CODER-CONVERGENCE (fix-until-green), DISCOVERY-EXHAUST (research-until-comprehensive), WORKER-WATCH / WORKER-CONVERGENCE, AUDITOR-REFINE, ENGINEER-PLAN-REFINE, and the orchestrator's FOCUS-LOOP — each specializing an existing composite, each with a hard --max cap and a measurable terminate-on predicate. New binding doctrine skills/shepherd/doctrines/loop-templates.md; commands/loop.md points operators at the catalog. No new halt codes (reuses the v6.0.9 circuit-breaker set).

Discovery waves compile to Dynamic Workflows (operator request)

All discovery fan-out now compiles like coder/audit fan-out instead of dispatching as inline Agent batches. doctrines/workflow-compile-down.md §V documents INTRO-COMBO-WAVE and DISCOVERY-COMBO-WAVE as compile targets (gate-free, parallel-safe → one Promise.all of discovery + auditor [+ worker] spawns); intro-combo-wave.md and discovery-combo-wave.md adopt the compile framing; pipeline.md gains the missing DISCOVERY-COMBO-WAVE taxonomy row. The compiler cmd_graph.sh was already role-agnostic (spawns_for_node expands any role mix) and verified end-to-end — the change is a clarifying comment plus a fixed latent bug where a node typed dynamic_workflow would not have matched the compiler's literal node-type key.

Spawn flow — per-sprint context certification, teammate Dynamic Workflows, default FOCUS-LOOP (operator request)

Four coordinated /shepherd:spawn fixes so the team substrate behaves as designed:

• Per-sprint context-certification wave. The spawn-flow walkthrough now makes the root's INTRO-COMBO-WAVE explicit (it was mandated in agents/shepherd.md but omitted from commands/spawn.md's flow): @discovery × N gather ground-truth, intro-mode @auditor × 2 certify it (regression / carry-forward / freshness) — the sprint's own certifiable current context. Always-on under spawn (every T-shirt) and fresh per sprint — each --scope patch/--auto sprint and each --parallel sibling certifies its own; a prior sprint's context is never inherited. intro-combo-wave.md gains the spawn framing.
• Teammate-conductors compile their lane fan-out. The contract required it (dispatch-cascade.md §IV-bis, conductor.md) but no operational instruction existed, so teammates dispatched in-context. Added the explicit shctx graph compile --segment=<entry> --verify → run → shctx graph mark sequence (in-context fallback only on confirmed runtime failure) to the commands/spawn.md teammate boot prompt and agents/conductor.md Step 2 + hard-prohibition #22; hand-rolled in-context step fan-out is a PRIMITIVE-INVERSION off-substrate violation. Reconciled the self-contradictory SKILL.md §X: under spawn, both root and each teammate compile their respective fan-out (mode-agnostic).
• Root adopts the FOCUS-LOOP by default on team init. Coordinate mode is reframed as operating the Pattern-6 FOCUS-LOOP (wake → act → probe, opened at SEED-VERIFY), entered the instant teammates spawn — the active engine, not a passive focus-record write backstopped only by coordinate_drive_guard.sh. The root stays engaged and drives until CLOSE-FINALIZE.
• Long-running conductors adopt their own FOCUS-LOOP. A teammate-conductor opens a lane-keyed focus loop at Step 0 (lane start, before any node — so a teammate that skips INTRO still gets one) and runs wake → act → probe over its lane micro-Stage-Graph, refreshing at each wave so a long lane doesn't drift.

All four default-on, config-gatable via the new [focus].loop_default key; doctrine framing in coordinate-active-drive.md.

Flock profile polish

Description-field shrink for the two genuinely bloated meta profiles — conductor (198→157 chars) and shepherd (195→152) — moving mode/tier detail into the body; planter/auditor tightened further. Frontmatter already consistent across all nine (name → color → model → thinking → description → tools); no value changes.

Foundation

• Version moved to 6.1.2 across the six sources of truth (plugin.json, marketplace.json ×2 keys, both SKILL.md frontmatters, README.md, this file).
• Removed a stray tracked err.txt and the dogfood repo's .artifacts/ tree reorganized onto the standard layout (plans/+reports/ → docs/).
• All new bash honors house style — set -uo pipefail, source _lib.sh || exit 0, exit-0-always hooks, resolve_namespace/resolve_workdir (never hardcoding .artifacts/.shepherd), and graceful-empty reads. New test auto-discovered by skills/context/tests/run.sh.

---

v6.1.1

What's Changed

• v6.1.1 by @FL03 in #139

Full Changelog: v6.1.0...v6.1.1

v6.1.0

What's Changed

• release: v6.1.0 by @github-actions[bot] in #137

Full Changelog: v6.0.9...v6.1.0