Changelog
August 6, 2026 · View on GitHub
Unreleased
Direction B sandbox
- fix: add
--writabletodelegate-codex.sh submit/resume— with no sandbox flag a job inheritscodex exec's own default, which is read-only unless the user setsandbox_modein~/.codex/config.toml, so implementation jobs finishedDONEhaving written nothing - fix: let an explicit
PARTNER_CODEX_BINwin over the parent job's binary onresume; inheriting it unconditionally made the env var silently inert, so a wrapper CLI never ran and the round was wasted - feat: record the effective sandbox (
read-only/workspace-write/codex-default) in each job'smetaand in--dry-runoutput - docs: state the real default in usage and in Direction B Phase 2, and add a Phase 3 check that
DONEplus an emptygit status --shortis an anomaly, not a completed task - test: add
tests/test_delegate_sandbox.pycovering the three sandbox states, flag exclusivity, the resume-coverride, and both resume binary-resolution paths
v2.0.1 (2026-07-29)
Bounded Claude planning
- feat: add
scripts/run-claude-plan.py, a config-only Claude planner with a validated 24,000-character evidence packet delivered over stdin, safe mode, zero tools/subagents, wall and no-event timeouts, and a Claude-CLI-enforced API budget - feat: persist bounded sanitized event envelopes, visible-text checkpoints, exact configured/observed model/session, runner and packet hashes, cost metadata, atomically created plans, and same-session recovery instructions under
.partner/ - fix: distinguish authentication, upstream idle, local idle timeout, wall timeout, budget stop, tool-use violation, and protocol error instead of treating every Fable failure as login trouble
- fix: invalidate verification when an identity changes; refuse unverified/non-Claude roles, malformed or secret-like packets, silent truncation, incomplete/conversational plan output, nonzero nominal success, observed model/session mismatches, output overwrite races, and all automatic model fallback
- fix: bound partial-line, visible-output, event-log, process-group termination, and expanded secret-redaction paths so watchdogs cannot be defeated by byte trickles, inherited pipes, or common bearer/password/GitLab/AWS/Google credentials
- refactor: share the nested Claude environment cleanup between setup smoke and the bounded planner
- test: add mocked stream/budget/timeout/config boundary coverage, repository behavior prompts, and the full unit suite to GitHub Actions
- docs: define who builds the bounded packet, its exact contract, the honest budget telemetry boundary, and the fixed recovery path
v2.0.0 (2026-07-29)
Setup, identity routing, and safety
- fix: strip host-injected Claude Code environment markers before spawning the real Claude CLI, preserving first-party OAuth instead of triggering a nested-session/login failure
- fix: restore delegate-codex.sh auto --skip-git-repo-check for non-git --repo targets, lost in the v1.5 rewrite (jobs against non-git dirs FAILed immediately)
- fix: align the model matrix rows and synchronize the matrix/settings header grid without removing the routing motion
- feat: rebuild the setup UI as a kinetic local Agent routing console with live role mapping, purposeful motion, responsive layout, and reduced-motion support
- feat: localhost single-page setup UI with a taste-skill guided decision rail, concrete model matrix, exact diff preview, preview-bound confirmation, and smoke test without repeated chat questions
- config: balanced preset fast_worker now uses the detected Codex model with high reasoning effort
- feat: identity matrix — three cross-vendor identities (deep_reasoner / fast_worker / arbiter), each with its own backend/model/effort; schema v2 with fail-closed v1 migration (
5a1f3d7,52ad950,da269ce) - feat: arbiter blind-solve protocol + 搭子,试跑 first-run tryout; goal.md task table drops owner in favor of identity (
f329daf) - feat: idea-king adds 分工 (Assignment) section to Partner work-split reviews (
18dd247) - fix: wire per-task role decision into the split flow; clarify owner vs role (
983457f,493d561, superseded by the identity matrix)
Protocol, receipts, and verification
- test: make the dual-host sandbox matrix self-contained with deterministic fake CLI binaries so CI does not depend on installed Claude or Codex tools
- test: dual-host CI sandbox matrix — install order, idempotence, fail-closed (
6333e8c) - fix: redirect codex exec stdin from /dev/null to prevent hung background jobs (
0d673cc) - docs: README bilingual rewrite — setup wizard, host self-ID, receipt v2, opt-in full protocol (
54ce055) - feat: goal-sync.py hash-checked goal.md read/write, no silent lost update (
94349df) - test: test-prompts.json +4 goal-to-pr case incl. ordinary-pr-no-trigger negative (
d99b630) - feat: Plan→Goal→PR→Verification protocol, references/goal-to-pr.md (
f5b9232) - feat: extend Partner Session Receipt with host/scope/config_source/roles_used, schema v2 (
5246d45) - feat: activate Sub Agent three-level routing, partner-* > user agent > generic Task (
3d5e077) - test: test-prompts.json +9 case — 4 setup, 4 host-adapter, 1 idea-king (
c6020f5) - feat: partner-setup.py wizard engine + references/setup.md, "搭子,配置" first-run setup (
5e3744b) - feat: install.sh --configure forwards to the terminal setup wizard (
9e95036) - feat: idea-king absorbs Occam/Murphy/Coase laws, ported from installed copy (
51a6389) - feat: idea-king clarify-to-95% pre-verdict protocol, headless degrades to Open Questions (
d86ae57) - refactor: split SKILL.md into Partner Core + references/codex-driven.md with host detection (
4cc5ccd) - feat: partner-config.py TOML-subset config engine (schema v1) with tests and docs (
760d938) - feat: delegate-codex.sh --role injection from partner-config, with --dry-run and tests (
63b6807) - feat: install.sh writes host= marker into .install-meta (
d9eab44) - test: run-test-prompts.py supports should_trigger:false negative cases (
b21cbe3) - fix: remove deprecated --enable web_search_cached from delegate-codex.sh (
30999a2)
v1.4.2 (2026-07-05)
- feat: idea-king absorbs official adversarial-review, grilling, and packet hygiene (
a1ddbc9) - docs: add idea-king adversarial-review showcase GIF to both READMEs (
4ab807c) - docs: add Red Skill submission copy for 搭子.skill (
9ebc3e7)
v1.4.1 (2026-07-04)
- feat: execution-channel routing + evidence-backed prompting principles (
ab679e8) - fix: harden bidirectional delegation utilities (
f5539bb)
v1.4.0 (2026-07-03)
- feat: bidirectional Partner + idea-king + frontier prompting & memory protocol (
a9b5dde)
v1.3.0 (2026-07-02)
- Release partner skill v1.3.0 (
e7266e8) - feat: showcase redesign with green palette, GSAP animation, and GIF (
e67966c) - fix: replace curly quotes with straight ASCII quotes in img tags (
764f4bb)
Earlier (2026-06-24 – 2026-07-01)
- Showcase and release polish: reproducible cost ledger, README parity release gate, README language split (
ad19e6c,efed0d8,8ebb5cf,3418b60,a266e94,b984911,c35955b,2eac3c1) - Public readiness: Partner session receipt, same-session Claude strategy, release preparation (
258779f,f87a17e,0ae1d81,7fa24da) - Origin: initial Claude Codex relay skill, renamed to Partner (
1ba62e6,5979fab)