§1 — Outcome

A fresh Codex or Claude session must automatically receive the applicable user-global and project-scoped rules, and must be able to prove what it loaded. “Read by default” means instruction files are in the initial effective chain and skill metadata is discoverable; full skill bodies remain trigger-loaded so context stays bounded.

Canonical user instructions feed agent-specific entrypoints, then project and target-path instructions, with a context manifest proving the loaded chain.
One canonical policy source feeds compatible entrypoints; a target-path preflight adds the applicable project chain and records evidence.
✓ landed 2026-07-29

§2 — Implementation

Added the reusable target-path context manifest and reckon agent-context doctor, then linked ~/.codex/AGENTS.md to the canonical policy and raised the measured project budget to 65,536 bytes. Live preflight passed with 32,782 bytes of headroom. Full record: §2–§3 landed (commits 02d7276, 32d7e46).

✓ landed 2026-07-29

§3 — Acceptance and evidence

Exercised 12 black-box context fixtures and the complete Python suite: 1,193 passed in 28.84 seconds. The live manifest reports no findings and records content-free path, byte-count, digest, skill-discovery and activation evidence. Full record: verified evidence.

§ Decisions

What is the canonical user policy source?

One canonical source prevents Claude and Codex policy drift while preserving each agent's discovery convention.

§ Followups

Implement the compatibility shim and context doctor

Build the cross-agent compatibility layer and prove it with black-box launch fixtures before changing any project instructions.

Project: reckon
Plan:    codex-agent-context-compatibility (http://localhost:8765/reckon/codex-agent-context-compatibility.html)
Section: full plan
Tier:    opus

Context
  Codex and Claude use different user entrypoints and nested-loading behavior.
  The workstation needs one canonical policy and a verifiable effective chain.

State to read
  codex-agent-context-audit.html, ~/.agents/AGENTS.md
  current Codex configuration and Ambix/Nova AGENTS.md files

Scope locks / constraints
  Preserve one canonical policy source. Do not mirror stale HPC skills.
  Produce research-before and evidence-after documents with source links.

Done-when
  1. Compatibility shim, doctor and black-box fixtures land
  2. tests still green
  3. followup written + this followup marked resolved

§2–§3 landed — commits 02d7276 and 32d7e46; 12 context fixtures and 1,193 total tests passed; live doctor exit 0 with 32,782 project-budget bytes remaining.

Consume the context manifest in bounded execution

The canonical entrypoint, measured budget and reusable target-path manifest are shipped and verified. The dependent execution hardening plan can now require this preflight before dispatch and persist the manifest in its execution ledger.
Project: reckon
Plan:    execution-orchestration-hardening (http://localhost:8765/reckon/execution-orchestration-hardening.html)
Section: context preflight integration
Capability: orchestrator-level

Context
  Codex context compatibility shipped with a reusable manifest builder, CLI doctor, 12 black-box fixtures and a live zero-finding report. Integrate that manifest into bounded dispatch only after the remaining S5 provenance substrate lands.

State to read
  GET /plan/reckon/codex-agent-context-compatibility
  GET /plan/reckon/execution-orchestration-hardening
  docs/archive/codex-agent-context-compatibility-s2-s3-landed.html

Locked decisions to honour
  canonical-source → agents-home-canonical

Open decisions to surface (do not resolve)
  none

Constraints
  Preserve content-free digests and paths; never persist instruction contents or secrets. Respect S6 scheduling and its provenance dependency.

Done-when
  1. bounded execution invokes target-path preflight before dispatch
  2. interruption and worker-failure fixtures retain the manifest
  3. tests stay green
  4. followup written into the plan and this followup marked resolved

§ Research

Codex Agent Context Audit