Skip to content

Releases: nail00749/opencode-agent

v0.12.4 — consistent agent contracts and Git workflows

Choose a tag to compare

@nail00749 nail00749 released this 07 Oct 09:34
3b5dee7

Fixed

  • Resolve Git readonly-file instructions conflicting with authorized repository operations. Normal requested staging, commit and push are allowed; force pushes, deletion refspecs and prune require exact-command approval. Remaining destructive denials and active-writer restrictions are retained.
  • Add DevOps direct native claim access and align writer ask/deny, coordinator and reviewer/security guidance.
  • Add all-17 agent contract tests, runtime/generated identity checks and isolated native Git/DevOps acceptance; regenerate agents and update knowledge.

Verification

679 tests, lint/typecheck/build, sync/package/release gates, 4 e2e tests and Linux/macOS CI PASS.
Both rebuilt OpenCode 2.0.24 isolated smokes PASS, including local bare push without approval, rejected force approval with no execution/retry, active-lease restrictions and DevOps claim/edit plus four denials. Cleanup PASS; no external push or production/global changes.
Deep review and Security APPROVE after remote-deletion/prune approval fix; metadata review APPROVE.

Limits

Explicit task intent remains model guidance. Wildcard permission rules are not exhaustive shell-safety enforcement; expanded-rule redundancy remains advisory. All-agent contract tests do not imply an individual live run of every agent. No dependency/compatibility or global installation changes.

PR: #17

v0.12.3 — direct native claim for strict writers

Choose a tag to compare

@nail00749 nail00749 released this 06 Oct 15:22
770f27a

Fixed

  • Expose native gvozd_claim directly with codemode:false while retaining exact permission and all lease guards.
  • Cartographer and Docs can claim leases without opening execute or adding MCP grants.
  • Update direct-call prompts and add host catalog, transport and native execution regression coverage.

Verified

651 tests, lint/typecheck/build, sync/package/release-tag gates, 4 e2e tests and Linux/macOS CI pass.
Actual OpenCode 2.0.24 with the rebuilt local plugin passes isolated native smoke: parent reservations, child claim/edit, strict writers without execute, readonly no-claim, and four rejections per writer (preclaim, wrong assignee, wrong parent, out-of-lease). Denied writes remain unchanged; final idle wait and cleanup pass.
Deep feature review and release metadata review: APPROVE.

No dependency/compatibility changes or global installation changes. Live smoke currently requires local macOS paths; production installation is not covered by the isolated result.

PR: #16

v0.12.2 — native claim permission fix

Choose a tag to compare

@nail00749 nail00749 released this 06 Oct 11:44
11a2f2a

Fixed

  • Explicitly bind the native gvozd.claim tool to the gvozd_claim permission action.
  • Give Docs only the exact claim permission under its deny-all policy; Cartographer already has the grant.
  • Add real-config and registration regressions preserving role, assignee, parent, unclaimed and out-of-lease guards.

No broad execute/MCP grants, dependency or compatibility changes.

Verification: 642 tests, lint/typecheck/build, sync/package/release gates, 4 package e2e tests and Linux/macOS CI passed. Deep security and release metadata reviews APPROVE.

Actual host discovery in fresh Cartographer/Docs sessions remains to be verified after installation.

PR: #15

v0.12.1 — MR evidence reporting fix

Choose a tag to compare

@nail00749 nail00749 released this 06 Oct 10:15
8478978

Fixed

  • Separate local Git inspection from CLI discovery and forge preflight calls.
  • Distinguish pending/rejected requests from executed CLI failures and unverified authentication.
  • Do not infer CLI/keyring authentication failure from anonymous HTTP login pages.
  • Preserve confirmed host/repository/MR identifiers in commands and suggestions.

Permissions, dependencies and OpenCode compatibility unchanged.

Verification: 638 tests, lint/typecheck, build, sync/package/release gates, 4 e2e tests and Linux/macOS CI passed. Prompt and metadata review APPROVE.

PR: #14

v0.12.0 — forge, CI and runner CLI skills

Choose a tag to compare

@nail00749 nail00749 released this 05 Oct 19:12
6c43388

Added

  • Built-in forge-workflow, ci-workflow and runner-workflow skills using git, glab and gh.
  • Secure managed installation through setup and project sync, with doctor/package checks.
  • Pipeline/job/runner diagnostics and explicitly confirmed supported operations.

Changed

  • Default agent guidance and permissions no longer depend on GitLab MCP. Existing user MCP configurations are not automatically removed.
  • Local users must provide installed and authenticated CLIs; skills are guidance, not permission enforcement.

Verification

634 unit tests + 4 package e2e tests, lint, typecheck, build, sync/package/release gates and Linux/macOS CI passed.

PR: #13

v0.11.4

Choose a tag to compare

@nail00749 nail00749 released this 05 Oct 13:52

Did-you-mean hints for unknown commands; per-command help

v0.11.3

Choose a tag to compare

@nail00749 nail00749 released this 05 Oct 12:50

Fix zsh completion registration (compdef)

v0.11.2

Choose a tag to compare

@nail00749 nail00749 released this 05 Oct 12:01

Shell tab-completion (closes #9); cartographer claim grant

v0.11.1

Choose a tag to compare

@nail00749 nail00749 released this 30 Sep 15:39

Manual model pool selection

Release 0.11.0

Choose a tag to compare

@nail00749 nail00749 released this 30 Sep 14:44
653933f

0.11.0

Added

  • Goal/extreme mode (the extreme coordinator): every round runs
    measure → improve → verify against a metric until plateau or limit,
    with family-scoped auto-approval, CLI transport through a file intent,
    a round log plus record/status reporting, and per-layer config
    limits.
  • Permissions tab bulk actions: Allow all approves every pending
    request at once and Reset all returns the tab to the agent policy,
    so a full grant or a clean reset no longer needs row-by-row toggling.
  • Mandatory code-review-excellence in the review prompts: both review
    tiers carry the shared review skill, and gvozd doctor reports a
    missing review skill instead of silently running without it.

Guards, stated exactly as implemented:

  • The grant is family-scoped: it covers only sessions in the goal
    family, never the whole server.
  • Never-escalate command families and file-lease enforcement stay
    denied even under an active goal.
  • Loop commands come only from the start input: agents run exactly the
    recorded measureCmd/verifyCmd and cannot invent replacements
    mid-loop.