Checkpoint Codex Work

Create a durable, immutable handoff note for current Codex work. Use this when the user asks to checkpoint, wrap up, hand off, remember the work state, or when the post-compaction SessionStart context requests the deliberate handoff.

When no project is pinned, discover projects with bm_project(command="list") and pass the selected project's external ID as project on note/search calls. With several addressable projects, these calls require an explicit scope even when a default is configured.

Gather

Read ~/.codex/basic-memory.json, then the nearest project .codex/basic-memory.json; project keys override user keys:

• primaryProject, default omitted
• captureFolder, default codex/<git top-level directory name>
• placementConventions, optional
• sessionProfile, default general
• repository, required when sessionProfile is coding

Apply the bm-writing skill before drafting the note.

Gather repo evidence:

• the original objective that started the thread and why it mattered
• the latest user intent, including corrections or scope changes that supersede

the original objective

• the approach taken and why it solves the problem
• the current system state and practical impact
• tradeoffs, sharp edges, useful simplifications, and intentionally parked work
• git status --short
• current branch
• repository root and current working directory
• current Git SHA
• current pull request number, title, URL, state, base, and head when one exists
• changed files you touched
• tests or checks actually run
• failures or skipped checks
• decisions made in this thread
• unresolved blockers
• next action
• current username, hostname, and timestamp
• host-provided session_id, agent, codex_turn_id, trigger, and model

values from the checkpoint request, when present

Use direct, read-only evidence for repository and pull-request state. Do not claim a test passed unless you ran it or the user supplied the result. Treat host-provided session metadata as opaque identity data. Preserve exact non-empty values; never infer or rewrite them. For older requests supplying only codex_session_id, use that exact value as session_id with agent: codex. Do not emit the legacy field on new checkpoints.

Write

A checkpoint is a durable handoff, not a status dump or commit-by-commit changelog. Tell the story for a human or agent returning later. Treat it as a snapshot plus pointers to authoritative artifacts, not a replacement for tasks, decisions, plans, issues, pull requests, commits, diffs, checked-in docs, or source files.

Every invocation creates a new checkpoint. Never edit, replace, or append to an earlier checkpoint, even when the topic is unchanged.

Use the title:

Codex checkpoint - <UTC YYYY-MM-DDTHH-MM-SSZ> - <short topic>

The UTC timestamp is part of the immutable checkpoint identity and avoids filename-unsafe colons. If bm_write reports a title collision, retry with the smallest available numeric suffix such as - 2. Never resolve a collision by modifying the existing note.

Call bm_write with project=<configured primaryProject>, overwrite=False, and output_format="json" on every attempt. When primaryProject is omitted, discover projects and pass the selected project external ID explicitly. The frontmatter project field is descriptive metadata and does not replace the tool's project argument. The explicit non-overwrite flag must win even when the user's write_note_overwrite_default setting is true. Only accept a successful result with action: created; treat the tool error ALREADY_EXISTS as the title collision above, and stop on any other action or error.

Write a note to Basic Memory. For the general profile:

• title: the timestamped checkpoint title above
• directory: configured captureFolder
• tags: ["codex", "checkpoint"]
• frontmatter:
• type: codex_session
• status: open
• project: <primaryProject if known>
• cwd: <current cwd>
• started: <current timestamp>
• username: <current username>
• hostname: <current hostname>
• capture: deliberate
• agent: codex
• session_id: <host-provided Codex session id>, when supplied
• codex_turn_id: <host-provided Codex turn id>, when supplied
• trigger: <host-provided checkpoint trigger>, when supplied
• model: <host-provided model slug>, when supplied

For the coding profile, write type: coding_session and use the same common frontmatter plus these schema-required fields:

• repository: <confirmed stable repository identifier>
• repo_root: <git rev-parse --show-toplevel>
• cwd: <current cwd>
• branch: <git rev-parse --abbrev-ref HEAD>
• git_sha: <git rev-parse HEAD>

When the current branch has a pull request, also add the typed optional fields pull_request_number, pull_request_title, pull_request_url, pull_request_state, pull_request_base, and pull_request_head. Resolve the pull request with a read-only GitHub query; omit those fields when no PR exists. Write the number as a quoted string, for example pull_request_number: "123", so exact metadata queries behave consistently across storage backends. Never infer or copy repository/PR identity only from conversation text. Stop if the required coding fields cannot be proven.

Link Checkpoints From The Same Chat

When session_id is available, pair it with agent: codex for same-chat identity:

1. Before writing, search the configured primaryProject for both

codex_session and coding_session notes with filters={"agent": "codex", "session_id": "<exact host-provided id>"}. Also search legacy notes with filters={"codex_session_id": "<exact host-provided id>"}.

1. Page through both searches, deduplicate, and select the newest earlier

checkpoint by its valid started timestamp. Read it directly from primaryProject. Confirm the exact agent/session_id pair, or the exact legacy codex_session_id when shared identity is absent. Reject conflicting agent/session fields; a bare session ID is never cross-agent identity.

1. Add - continues [[Exact previous checkpoint title]] under ## Relations.

Do not edit the previous immutable checkpoint to add a forward edge; Basic Memory backlinks make the chain navigable in both directions. If there is no verified earlier match, omit the lineage relation. Never infer same-chat lineage from repository, branch, topic, timestamps alone, or lifecycle envelope notes.

Begin the body with # <exact note title>.

Use these sections, omitting optional ones that add no value:

• ## Summary: one concrete sentence that does not merely repeat the title
• ## Story: original objective -> latest user intent -> approach -> current

state and impact in substantive prose

• ## Working State: separate durable state from machine-local or fragile state
• ## Changed Files, when paths are useful for resuming
• ## Verification, for checks actually run and their outcomes
• ## References, for verified repository, commit, pull-request, issue, spec,

or documentation links

• ## Observations
• ## Relations, when the thread has an obvious graph target

Prefer repository-relative paths in the body. Required absolute repo_root and cwd frontmatter remain machine-local evidence. Label dirty or untracked files, ignored files, active processes, dev servers, temporary directories, and local tool caches as machine-local or fragile when they matter to resumption. Do not present them as durable project state.

Make the note pointer-first:

• name authoritative artifacts and include their stable identifiers or links
• summarize only the context needed to understand why each pointer matters
• use a relation for an existing graph note and a normal link or repository

path for artifacts outside the graph

• do not copy large plans, diffs, logs, or source files into the checkpoint

For GitHub-backed repository work, resolve the canonical repository URL with a read-only GitHub query. Render the current repository, current pushed commit, pull request, and any materially relevant GitHub issues or commits as Markdown links under ## References and where they appear in prose. Use the canonical URL returned by GitHub for pull requests and issues. Before linking a commit, verify that GitHub can resolve that SHA in the confirmed repository. If a commit is local or unpushed, keep the SHA as code, label it local or unpushed, and do not construct a GitHub link that may not exist. Do not turn an ambiguous bare issue number or SHA into a link without proving its repository.

Use observations to distill durable facts for structured recall rather than duplicating every narrative sentence:

• [result] for concrete outcomes
• [decision] for each decision made or preserved
• [blocker] for each unresolved blocker
• [next_step] for the one primary next action; include exactly one
• [verification] or [changed_file] only when the item is itself important

project memory, not merely supporting detail

Do not create separate Decisions, Blockers, or Next Action sections with plain bullets. Omit empty categories instead of writing placeholder text such as "None."

Relations are not observations. Put them under ## Relations using Basic Memory relation syntax, for example - relates_to [[Exact existing note title]]. Never write [relates_to] or a bare memory:// URL as an observation. Only add a relation when its target is an existing checkpoint, task, decision, spec, issue, or PR note. The verified same-chat continues edge is the checkpoint lineage relation; do not add a second generic relation to that same target.

Confirm

Reply with:

1. one sentence summarizing what the checkpoint preserves
2. the exact resume identifier selected from the successful JSON result
3. the one primary next action
4. exactly one fenced resume command as the final block:

$bm-orient "<exact returned resume identifier>"

Choose the first non-empty returned value in this order: permalink, file_path, then title. Use the returned value verbatim; never construct or guess a permalink or file path.