v2.5.0 — plan-mode read-only auto-approve, context economy, zero-idle governor

Tale Mode is a discipline plugin for Claude Code and OpenAI Codex — it makes any Claude work like a careful senior engineer (verify the real code, source every decision, run an independent review before "done"), and ships a self-armed autonomous loop that keeps going until a real check passes.

Highlights since v2.3.0

Plan-mode read-only auto-approve (the 2.5.0 headline, #14). Tale-mode's plan/kickoff ceremonies live in plan mode, where every shell command raises a permission dialog — turning careful investigation into a wall of prompts. A new fail-closed PreToolUse hook auto-approves a Bash call only when it can prove the whole command read-only; everything else silently falls through to the normal dialog, so it can never block or approve anything you wouldn't have. The classifier is default-deny and adversarial: it rejects $/backticks/backslashes/process-substitution/heredocs outright, rejects every output redirect except a token-boundary >/dev/null, catches quote-obfuscated flags by matching on dequoted tokens, and positively guards every whitelisted binary that has a write/exec flag (find/sort/git/sed/jq/…). Review-hardened: git config approves named-key reads only — the dump forms can spill credential-bearing config into model context. Kill switch: TALE_PLAN_APPROVE=0. The honest residual-risk write-up is in SECURITY.md; the 118-check test suite is deliberately fail-case-heavy, because a guard that can't fail isn't a guard.

Context-economy discipline (2.4.0). The skill now treats the context window as a budget: fan bulk work out via Workflow/subagents, demand terse conclusions back instead of raw dumps, and batch round-trips — so long sessions stop drowning in their own transcript.

Governor v2.0.0 — zero idle cost (#13). The optional Layer-2 governor now costs nothing while things go well: exactly one read-only model escalation per stuck goal, on both hosts. Install separately: /plugin install tale-mode-governor@tale-mode.

The map. New ARCHITECTURE.md documenting the three-layer model (discipline skill → deterministic loop → advisory governor), loop mechanics, and the trust model.

Install

/plugin marketplace add alicicek/tale-mode
/plugin install tale-mode@tale-mode               # core (free)
/plugin install tale-mode-governor@tale-mode      # optional governor

Requires Claude Code ≥ v2.1.154 (kickoff auto-arm verified on ≥ v2.1.195). See the README for the full tour and SECURITY.md for the threat model.

Verification

Every hook path is covered by a deterministic gate — bash tests/verify-cross-platform.sh runs the six unit suites (164 / 29 / 10 / 65 / 118 checks + 36 for the governor), claude plugin validate, and Codex parse/engagement checks. The plugin is dogfooded: changes to it are built with its own method and reviewed by the plan-reviewer agent it ships.

v2.3.0 — cross-platform, helper skills, and the Codex governor

Tale Mode is a discipline plugin for Claude Code and OpenAI Codex — it makes any Claude work like a careful senior engineer (verify the real code, source every decision, run an independent review before "done"), and ships a self-armed autonomous loop that keeps going until a real check passes.

This is the first tagged release. It captures everything from the v2 line through today: the plugin runs on both Claude Code and Codex, arms itself from committed gates during a kickoff phase, and now has an optional read-only governor on both hosts.

Highlights since the v2 baseline

Cross-platform (Claude Code + Codex). The skill and the autonomous loop run on both runtimes. On Codex the phase workflows surface as skills, and a one-time touch ~/.tale-mode-allow-cwd-root opt-in lets the loop find your project root. Verified against a live Codex session, not just documented.

Helper skills. trust (documents the two user-only opt-in files and the security model), seed-gates (suggests committed gates from your package.json/CI — suggest-only, never auto-writes), and end-phase (the explicit off-switch for phase enforcement). All tuned to load only when you ask.

Autonomous loop, more capable.

• Committed-gate auto-arm during a /tale-mode:kickoff-phase phase, gated on a content-hash you trust.
• Optional no-progress stop (TALE_NO_PROGRESS_N): disarm cleanly after N identical failing rounds instead of grinding to max_rounds.
• Multi-line gates (a gate value may span lines and runs as one script).

Layer-2 governor (optional companion, v1.2.0). Once the loop is stuck, a read-only fresh-context reviewer names the thing the deterministic gate can't see. On Claude Code it's a Sonnet agent hook; on Codex it's a new hook that spawns one OS-sandboxed codex exec --sandbox read-only reviewer per stuck goal, guarded against recursion. Advisory only — it never blocks. Install separately: /plugin install tale-mode-governor@tale-mode.

Install

/plugin marketplace add alicicek/tale-mode
/plugin install tale-mode@tale-mode               # core (free)
/plugin install tale-mode-governor@tale-mode      # optional governor

Requires Claude Code ≥ v2.1.154 (kickoff auto-arm verified on ≥ v2.1.195). See the README for the full tour and SECURITY.md for the threat model.

Verification

Every hook path is covered by a deterministic gate — bash tests/verify-cross-platform.sh runs the four unit suites (164 / 29 / 10 / 65 checks + 26 for the governor), claude plugin validate, and Codex parse/engagement checks. The plugin is dogfooded: changes to it are built with its own method and reviewed by the plan-reviewer agent it ships.