fix(plugins): author Codex checkpoints after compaction

[phernandez]

Why

Codex lifecycle capture was treating harness telemetry as durable knowledge. bm hook flush
projected sparse SessionStart/PreCompact envelopes into synthetic session and tool_ledger
notes, which produced low-value session garbage instead of useful handoffs.

The attempted automatic checkpoint path had a second mismatch: Codex ignores PreCompact
stdout, and the transcript adapter expected a Claude-shaped transcript rather than Codex
response_item events. Hooks could be enabled and visibly capture events without the active
agent ever authoring an actual checkpoint.

The Codex marketplace instructions also made it easy to sparse-checkout only plugins/codex,
which omits the repository marketplace manifest and causes the "marketplace root does not
contain a supported manifest" installation error.

What Changed

• Keep lifecycle envelopes as owner-private local trace. basic-memory hook flush now archives
  valid envelopes locally and never writes knowledge-graph notes.
• Add a Codex PreCompact-to-Stop handshake. PreCompact records a private checkpoint request;
  Stop blocks once and asks the active model to run codex:bm-checkpoint from its compacted
  working context.
• Parse real Codex response_item.payload transcript records, including input_text and
  output_text, for checkpoint context.
• Apply accumulated user/project redactKeys and redactPaths to every authored checkpoint
  field, failing closed instead of writing when the privacy gate cannot be satisfied.
• Remove projected session / tool_ledger recall from the Claude and Codex skills and update
  hook status language to report archived trace and pending checkpoint requests.
• Document the correct marketplace install shape: leave sparse paths empty, or include both
  .agents/plugins and plugins/codex; the marketplace manifest should not move.
• Register and validate the new Codex Stop launcher and include it in release pinning.

Implementation Details

• src/basic_memory/hooks/archive.py owns the trace-only flush path. Invalid envelopes remain
  pending for inspection; valid envelopes move into the private processed archive under the
  existing retention and locking rules.
• src/basic_memory/hooks/checkpoint_requests.py stores one atomic, owner-only request per Codex
  session. The first Stop emits a decision: block response; the re-entrant Stop clears the
  request and continues when stop_hook_active is true, preventing a loop.
• The active Codex model remains the checkpoint author. No nested agent or telemetry-to-note
  fallback manufactures memory from lifecycle payloads.
• bm-checkpoint now scrubs title, directory, tags, frontmatter, body, repo/PR evidence, and
  embedded paths before write_note; required path fields use [REDACTED_PATH] instead of
  leaking or disappearing.
• src/basic_memory/hooks/projector.py remains as a compatibility re-export while the graph
  projection implementation is removed.
• The three PEP 723 launchers are pinned to implementation commit
  f802a01ddd0d6eab5b83cc7c684619776b7df1db, which already contains the Stop command. Keeping
  the pin in a second commit avoids a wrapper/runtime circular dependency.

Testing

Automated

• uv run pytest -q --no-cov tests/hooks tests/cli/test_hook_command.py tests/cli/test_coding_session_context.py plugins/codex/hooks tests/test_codex_plugin_package.py: 238 passed
• just package-check-codex: passed against the final hook pin
• just fast-check: passed
• just package-check: passed across Claude Code, Codex, shared skills, Hermes, and OpenClaw packages
• just doctor: passed
• uv run python /Users/phernandez/.codex/skills/.system/plugin-creator/scripts/validate_plugin.py plugins/codex: passed
• git diff --check: passed

Manual

• Ran a local Codex lifecycle smoke flow: PreCompact created one checkpoint request, the first
  Stop returned the one-time block instruction, and the re-entrant Stop returned
  {"continue": true}.
• Flushed seven existing and two synthetic envelopes. All moved to local archive; pending
  envelopes and checkpoint requests returned to zero.
• Queried session and tool_ledger notes after the exact flush timestamp: 0 results, proving
  the flush path no longer promotes telemetry into the graph.

Risks / Follow-ups

• The Stop hook can request an authored checkpoint but cannot guarantee that a model follows the
  skill correctly. The loop guard fails open, the privacy policy fails closed before writing,
  and neither path falls back to synthetic telemetry notes.
• Existing synthetic session/tool-ledger notes are left untouched; deleting user data should be
  a separate, explicit cleanup operation.
• Existing remote marketplace installs receive this behavior only after this change is merged
  and the plugin is refreshed in a new Codex thread.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 54d5813ce0

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 2c9a3217c5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".