v0.5.0 — Claude Code & Codex plugins

Install coding-posture as a plugin and keep it updated in place. No change to the skill itself.

Claude Code

/plugin marketplace add alexei-led/coding-posture
/plugin install coding-posture@coding-posture

Update: /plugin marketplace update coding-posture then /reload-plugins (or enable auto-update in the /plugin menu).

Codex CLI

Clone the repo, then /plugins → add the local marketplace → install coding-posture; or point ~/.codex/config.json skills at skills/coding-posture. Update with git pull.

Details

• Repo root doubles as marketplace + plugin for each tool (.claude-plugin/, .codex-plugin/, .agents/plugins/), all sharing the single skills/coding-posture/ — no duplication.
• Still works as a plain SKILL.md skill for Hermes, Pi, Cursor, and any AGENTS.md-compatible agent.

v0.4.3 — docs, rename, repo hygiene

Docs and repo housekeeping. The skill (SKILL.md) and the eval are unchanged from v0.4.2.

• Renamed the repo and package to coding-posture — it's a skill, not an engine. (Old URLs redirect.)
• README rewrite: added a validated "How it works — theory and evidence" section (real citations, evidence-strength labels), restructured reader-first with a usage diagram and scannable tables, and reframed the project as something to use and try rather than a tentative MVP.
• Repo hygiene: track gitleaks pre-commit/pre-push hooks (scripts/git-hooks/); ignore .ruff_cache/. Enable locally with git config --local core.hooksPath scripts/git-hooks.

Try it: drop skills/coding-posture/ into your agent, or run the behavioral eval in eval/ (85% vs 70% without the skill, +15pp).

v0.4.2 — theory & evidence in README

Docs-only patch. SKILL.md and the eval are unchanged from v0.4.1.

Adds a "How it works — theory and evidence" section to the README, validated against the literature (via perplexity) with each claim labelled by evidence strength:

• Mechanism — procedures work via in-context conditioning; induction/iteration heads and CoT as a plausible substrate (Wei 2022, Olsson 2022, ICML/NeurIPS 2024). Grounded substrate, not a complete theory.
• Procedures, not personas — a supported guideline (Zheng EMNLP 2024), not a universal law.
• Self-selection — holds for strong/calibrated models (Route-to-Reason), not categorically better than a router.
• What the checklists encode — highest-evidence levers (self-debug, self-correction degrades, context-before-editing, verifier gaming); no clean RCT, hence our own eval.
• Our eval — +15pp, one run/one model: directional, not a settled effect.
• Explicitly debunks the unsourced "~150–200 instructions" figure.

v0.4.1 — eval refinements (+15pp validated)

Patch: eval/test refinements. SKILL.md is unchanged from v0.4.0.

Changes

• The eval shim auto-loads a local .env (gitignored), so runs are self-contained.
• Fixed unfair assertions in the treat-urgent-auth eval case. Its prompt gives no code, so the disciplined model correctly asks for it first — the old assertions demanded it "verify the fix," which was unsatisfiable. Now: treat auth as high-risk, ask before blind-patching, commit to verify/check-regressions before declaring done, no fake success.

Eval result (rerun, baseline mode, gpt-5.4-mini)

• with-skill 85% vs without-skill 70% → +15pp (per-assertion pass rate).
• The treat-urgent-auth case now passes 4/4 with-skill vs 2/4 without — the fix is validated and the skill shows clear lift there.
• Caveat unchanged: N=1 per case, so this is directional. Run eval/run-skill-evals.sh with more iterations for steadier numbers.

v0.4.0 — always-on snippet + behavioral eval

Adds the two highest-leverage follow-ups from the architecture review. The shipped SKILL.md is unchanged.

always-on snippet

always-on-snippet.md condenses the always-true invariants — verify by running, never fake green, no destructive commands without scope, and a pick-a-mode pointer — for pasting into CLAUDE.md / AGENTS.md. These belong always-on because a conditionally-loaded skill can fail to activate, and they're the rules you want on every turn.

behavioral eval

eval/ measures whether the skill changes behavior, reusing the agent-skills-eval runner (LLM judge + with/without-skill baseline) from cc-thingz rather than a bespoke harness. Five cases target the skill's delta over agent defaults — gaming an impossible test, thrashing after repeated failures, optimizing before measuring, migrating without rollback, trivializing an urgent auth patch.

OPENAI_API_KEY=sk-... bash eval/run-skill-evals.sh

Verified locally: eval schema, the layout assembly, and that every flag matches the agent-skills-eval CLI. The paid run (and its numbers) is left to the maintainer.

Honest status

Still unproven on numbers — but now testable. That was the gap; this closes the tooling half of it.

v0.3.0 — more modes + instruction polish

Adds three research-grounded modes and an instruction-quality pass.

New modes

• optimize — measure first; baseline before/after; don't optimize what merely looks expensive.
• upgrade — dependency/version bumps: read breaking changes, account for transitive deps, no blind search-replace.
• integrate — external APIs/services: read the contract, handle error paths (auth, timeouts, retries, rate limits, partial responses). Tool/API misuse is the most-documented agentic failure (MAST, Microsoft taxonomy).
• Folded: concurrency checks into review; safe code-removal into refactor.

Instruction-quality pass

Reviewed against an AI-instruction rubric: removed italics, dropped the human-facing rationale section (signal density), added a one-line output contract (Mode: <name> — <reason>), added a non-trigger to the description, and covered the can't-run-the-check case in the verify invariant.

Honest framing (README "Where this fits")

A conditionally-loaded skill is the right home for a growing catalog of task modes — but the always-true invariants (verify by running, never fake green, no destructive commands) are most reliable in always-on instructions (CLAUDE.md/AGENTS.md), since a skill can fail to activate. The skill should be the delta over what a frontier agent already does by default.

Status

Still unmeasured. Per Anthropic's evaluation-driven guidance, the real next step is an eval that shows these modes change behavior — not more modes.

v0.2.0 — rebuilt as a SKILL.md skill

Breaking rebuild. The keyword-scoring engine is gone; the product is now a single self-selected SKILL.md skill.

What changed

• Removed the Python scoring engine, the duplicate TypeScript port, the CLI, the Hermes plugin code, postures.json, and the keyword gate.
• Added one artifact: skills/coding-posture/SKILL.md. The agent reads the modes and picks one from context — no router.
• Modes renamed to plain task names (debug, fix, review, test-first, refactor, migrate, spike, unstuck); persona-style IDs and unused axes dropped.

Why (grounded in research)

• Naming a persona does not reliably change model behavior; specifying a procedure does. (Zheng et al., EMNLP 2024)
• Context-based self-selection beats keyword routing on strong models. (Route-to-Reason, 2025)
• Modes encode the practices with the strongest evidence: execution-feedback loops (self-debug), verifying by running rather than re-reading (self-correction caveat), gathering context before patching (behavioral drivers), precise localization, small diffs, clarifying underspecified requirements (ClarifyGPT), and not gaming the tests (verifier gaming).

Install

Works unmodified in any SKILL.md-compatible agent (Hermes, Claude Code, Codex, Cursor, Pi). See the README.

Status

MVP. The bet — that these procedural checklists improve agent behavior — is not yet measured. No eval here compares behavior with and without the skill.

v0.1.2

Fix Hermes directory-plugin installs from ~/.hermes/plugins by adding the plugin root to sys.path before importing the bundled Python package. Also keeps v0.1.1 posture-selection and Pi-context improvements.

v0.1.1

Fixes follow-up review items: avoid substring false positives, use a single JSON posture catalog for Python/Pi, switch Pi injection to ephemeral context messages, rename Pi prompt template to avoid /posture conflict, and add ops-runner posture.

v0.1.0

Initial MVP: Hermes plugin, Pi package extension/skill/prompt, CLI wrappers for Claude Code, Codex, and Pi.