Skip to content

Releases: phenomenoner/Chatgpt-Codex-App-Plus

Context Canvas Codex v0.7.0 — bounded reflection preview

Choose a tag to compare

Context Canvas Codex 0.7.0

This release adds the opt-in context-canvas-reflection v0 preview skill while
leaving the Context Canvas 0.6 runtime, MCP server, hooks, persistence, references,
and snapshot behavior unchanged.

What is new

  • one evidence-first reflection pass at a meaningful trajectory checkpoint;
  • a bounded CONTINUE, INVESTIGATE, or ESCALATE disposition contract;
  • a closed optional Canvas read allowance and provenance-labelled proposal-only
    writeback;
  • owner-override retirement, shared per-task budgets, stop conditions, and
    experimental utility/retirement guidance;
  • bilingual product documentation and focused static contract coverage.

Implicit invocation is disabled in the public plugin by default. Invoke
$context-canvas-reflection explicitly or opt in with a bounded trigger policy in
a user-controlled AGENTS.md. The skill is advisory: it does not own replanning,
rollback, pause, publication, delivery, external effects, or Canvas lifecycle.

Install

codex plugin marketplace add phenomenoner/Chatgpt-Codex-App-Plus --ref context-canvas-codex-v0.7.0
codex plugin add context-canvas-codex@codex-app-plus

Restart Codex and open a fresh task before checking pickup. Plugin catalog,
configuration, or cache visibility alone does not prove that a running task loaded
the new skill or hooks.

Verification

  • root suite: 9 tests passed;
  • Context Canvas suite: 88 tests passed, 1 skipped;
  • skill structure validation passed;
  • public manifest/lock validation passed with zero errors;
  • git diff --check passed.

Context Canvas Codex 0.6.0

Choose a tag to compare

Context Canvas Codex 0.6.0

Context Canvas Codex 0.6 makes Canvas an optional session-navigation and explicit long-context offload layer. It separates hook transport identity, semantic task maps, retrievable references, and historical tool snapshots; missing Canvas state never blocks the underlying work.

Highlights

  • Add explicit policy-redacted text references with bounded put, search, read, preview, and idempotent delete operations.
  • Return digest-bound UTF-8 match ranges and directly usable bounded read hints for content-search hits.
  • Add deterministic log-v1 and strict line-oriented search-results-v1 preview lenses. Preview segments are exact stored-source byte slices, remain within the caller's budget, and never rewrite or replace the stored body.
  • Keep PostToolUse snapshot persistence off by default. snapshot_capture_next requires one nonempty exact tool name, rejects omission, empty input, and wildcard scope before mutation, and arms one expiring request that is consumed once.
  • Make interrupted reference-pair creation retry-convergent after exact body or manifest verification, while surfacing unverified one-sided state as bounded corruption without deleting it.
  • Store generated snapshot-blob provenance as path-bound metadata so literal marker-shaped caller JSON round-trips unchanged through bounded read, export, pin verification, and garbage collection.
  • Migrate exact manifest-bound legacy hook generations in place, preserve peer handlers, and refuse duplicate, ambiguous, drifted, or foreign ownership. Uninstall now requires a canonical supported manifest or the exact current adapter before deleting any managed group.
  • Keep Canvas outside authorization, source-of-truth, WAL, release-gate, and workflow-engine roles. Hook IDs remain transport provenance only.

Compatibility and boundaries

  • Plugin package: 0.6.0+codex.20260818075059; MCP server: 0.6.0.
  • Existing v1 and v2 maps remain readable; their next intentional mutation persists the compatible v3 form.
  • Markerless V1 snapshot objects remain readable. V1 objects containing the retired structurally ambiguous blob marker fail closed rather than guessing provenance.
  • References and snapshots remain historical untrusted data and require live-source revalidation for current claims.
  • Snapshot opaque media remains explicitly uninspected. This release does not add transparent proxying, implicit retrieval, automatic lossy compression, default-on capture, or automatic task-map creation.
  • Python 3.11+ and Codex plugin support are required. Newly installed hooks may require explicit review in /hooks, followed by a fresh Codex App task.

Install

codex plugin marketplace add phenomenoner/Chatgpt-Codex-App-Plus --ref context-canvas-codex-v0.6.0
codex plugin add context-canvas-codex@codex-app-plus

Verification

  • GitHub Actions passed the public bundle checks on both Windows and Ubuntu for commit 34644f0af2e5657c6acd8a81055309a3110330b2.
  • The Context Canvas component suite passed 88/88 tests from source and 88/88 again from the exact installed cache; the public repository contract suite passed 8/8.
  • Source, installed cache, and the generated public lock matched across all 15 plugin files.
  • Independent release review completed without findings and preserved the separate source, installed-instance, native-host, and publication boundaries.
  • A fresh Codex App process loaded the exact installed plugin generation. Native acceptance verified rejection of a missing capture tool name before mutation, digest-bound search ranges, bounded reads, deterministic exact-slice preview, idempotent reference deletion, and zero automatic task-map creation.
  • The deterministic preview selected the same 217-byte source segment on repeat reads from a 4,403-byte log while omitting 4,186 bytes under a 1,024-byte response budget; both disposable references were deleted after verification.

The repository is community-maintained and is not an official OpenAI release.