Clarify in CLAUDE.md that the no-coupling rule only forbids file-level
dependencies between skills; mentioning or routing to another skill by
name remains allowed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add item_progression: parallel as an owner-gated scheduling extension:
the main workflow stays the sole orchestrator and cursor writer, workers
execute dependency-independent items in host-provided isolated
workspaces, and integration is serialized without advancing mainline
history until per-item re-verification passes. Ships the parallel
execution protocol reference (wavefront scheduling, contextPacket
delegation, active_items recovery records with branched state
transitions, item-local vs epic-global blockers, capability fallback),
mirrors the contract across bilingual WORKFLOW/CATALOG, ADR-005,
AGENTS.md and CLAUDE.md, and anchors it in the contract tests.
Reviewed by an independent heterogeneous reviewer over three rounds
(final verdict: approve; target sha256 a78d424e1d3a).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Implements LEARN-3 with frozen paired fixtures, crash-safe journals, target isolation probes, and measured learning-transfer verdicts.
Supersedes local checkpoints 3022df5, 715c831, 9fb5d0f, and fe635f1; those attestations were invalidated and never used for real model runs.
Owner direction: the README speaks to a first-time user of today's
CodeStable, not to its history or internal methodology.
- Removed the 'build -> evaluate loop' chapter (maintainer methodology;
lives in the repo, not the front page) and its tagline pointer
- v1 is now mentioned exactly once per language — inside the Upgrade
section, which also gained a line assuring that legacy knowledge
(compound/, features/, old tools) stays covered by v2 retrieval
- Design / catalog / workflow / roadmap sections rewritten to describe
what CodeStable is, not what it used to be; alias line reduced to
'cs-code-review is an alias of cs-review'
Doc contract suite stays green (96 passed).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rebuilt the three v2 content sections as the main narrative instead of
v1-era framing with patches:
- 'Design: thin harness, thick context' — write responsibilities not
steps; the hard-gate table per flow (incl. 3-round review cap, design
persistence, test-first); on-demand references for engineering
judgment; the six homes of knowledge with graduate-before-delete
- Skill catalog refreshed (cs-review three modes, epic inline-first
sub-designs, the single cs-code-review alias called out)
- 'Workflow and project memory' — one execution mainline, per-request
risk judgment, zero artifacts for ordinary work, retrieval-with-source
discipline, zero-migration v1 boundary
- Tagline gains the one-line methodology; roadmap logs the dogfood loop
Personal narrative sections (origin, orchestration comparison,
philosophy, build->evaluate loop) preserved as-is. Doc contract tests
stay green (96 passed).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Dogfood question: what happens when a work doc is ready to graduate but
the destination does not exist (no project docs convention, no ADR dir,
no requirements home)? Previously unspecified — agents would invent
directories, mis-file facts into lessons, or delete without graduating.
Three-tier default: CodeStable-owned targets (attention.md, lessons/) are
created on demand; project-side homes are never invented — the graduation
list proposes a destination for the owner to approve, and the work doc is
kept until approved; genuinely experiential content may land in lessons
(scope tagged 待归位) so nothing is lost. 96 passed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Dogfood feedback: an agent reported 'work docs deleted, only attention.md
kept' with no account of where the knowledge went — compress-then-delete
was required but unverifiable. Now the final report must list the
graduation destinations (which conclusion went into which project doc,
which lesson was distilled, or an explicit 'nothing to graduate') before
any work doc is deleted; no list, no deletion. cs-issue/cs-refactor also
gained the previously missing completion-disposal rule. 96 passed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Dogfood feedback: work/ entries like paseo-native-session-bridge.md give
no clue whether they are a feature, issue, or epic when it is time to
compress them. Two rules, zero new mechanism:
- work docs are named with a type prefix (feat-/issue-/refactor-/epic-),
so ls classifies and wrap-up knows the destination per type
- epic sub-designs are inline-first: a few lines under the item entry;
only risk-escalated items get their own work/feat-{slug}.md, carrying
'epic: {slug}' frontmatter with a backlink from the epic item row —
flat files with two-way pointers, no subdirectories. Epic wrap-up now
also sweeps its sub-item work docs.
Docs (WORKFLOW/README zh+en) synced. 96 passed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The skill has outgrown its name — it reviews designs and runs repo audits,
not only code. cs-review is the canonical entry; cs-code-review ships as
the one thin forwarding alias (the explicit exception to the no-shim rule,
kept for the carried-over v1 habit). References in cs / cs-feat / cs-epic /
cs-issue / cs-refactor, the Codex agent config, both contract-test skill
sets (V2_SKILLS + SHIM_SKILLS), and all public docs / CHANGELOG / ADR-004
updated. Also refreshed the leftover v1 openai.yaml wording. 96 passed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Owner decision after first dogfood hit of the cap. Escalation semantics
unchanged: exceeding the cap still surfaces both sides' reasoning for
human arbitration and never counts as a pass.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Dogfood feedback (session 001c553b): calling cs with a concrete request
recommended an entry and then waited for the user to re-invoke it — one
confirmation too many. cs now classifies intent: clear action requests
announce the chosen entry in one line and continue executing under that
skill's discipline in the same turn; advice requests still get a
recommendation only; ambiguous requests that would materially change what
happens next get exactly one focused question. Handoff does not widen
authority — the target skill's gates and checkpoints apply unchanged.
Safety anchors and docs (SKILL_CATALOG / WORKFLOW / README zh+en) synced.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Owner decision (applied from his working-tree edit): shipped SKILL.md files
carry no self-test metadata. Protection is preserved, not dropped — hard-gate
anchors move into tests/test_skill_contracts.py as direct body assertions
(SHIPPED_HARD_GATE_ANCHORS), plus an explicit no-contracts-frontmatter check.
The not-grep validator regression now samples build-cs-skill. CLAUDE.md rule
updated accordingly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Owner practice decision: on risk escalation the proposal is written into
.codestable/work/{slug}.md before independent design review and user
confirmation — the design trail is always on disk for high-risk work.
Contract anchor added; wrap-up section reconciled (high-risk work docs
are created at design alignment, others on cross-session/handoff need).
Sync WORKFLOW / README (zh+en).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Bring in the owner's 5-file thin-harness rework (SKILL.md, spec-standard,
quality-gates with Thin Harness / Context Plan / Evolution Compression
gates, fixture-patterns, openai.yaml) — contracts test now passes,
xfail marker removed
- Fix v1 contract residue in step 7: no more .codestable/reference/
sharing or cs-onboard/tools entry points; v2 skills invoke no runtime
tools, shared discipline is inlined per skill
- Trim contracts 13 -> 8: drop section-existence anchors already locked
by tests, keep behavioral anchors; add 'Additive-only evolution is a
failure mode' as the evolution anchor
- Mark ToolBackedWorkflow / ShimSkill as shapes with no v2 instance
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- work doc gains a status-and-open-items section so cross-session resume
has a defined landing spot for pending confirmations (recovery contract)
- cs-code-review: reviewer dispatch requires a self-contained context
packet (intent, scope, exclusions, expected graded findings); main flow
owns integration and final verdict (collaboration contract)
- cs-onboard: gitignored .codestable stops with a report instead of
silently proceeding (failure behavior)
- Sync WORKFLOW / README (zh+en)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- cs-code-review: add design review mode; cap blocking fix-and-rereview at
2 rounds, then escalate to human arbitration with the disagreement
- cs-feat: risk-escalated designs pass independent design review before user
confirmation; test-first when a test setup exists; completed changes get
independent review by default (trivial edits may skip)
- cs-epic: decomposition passes independent design review before user confirmation
- cs-issue / cs-refactor: independent review by default, trivial changes may skip
- Sync WORKFLOW / SKILL_CATALOG / README (zh+en)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Wyatt Fang