Save Session

save-session — tactical checkpoint

Self-validate after edits. Any change to this skill's files (SKILL.md, scripts/, references/, templates/, assets/) must be followed by ./scripts/validate.sh from the skill directory. Hard findings → create-skill Optimize lane.

This skill exists to preserve the transient context that durable docs usually omit: what the operator was actively doing, what should happen first next time, what was already verified, and which small set of local files or commits matter at handoff.

When the session exposed preventable friction, use session-introspection before writing the checkpoint. Save the highest-value prevention control in the checkpoint so the next agent can continue from the control decision instead of rediscovering the same inefficiency.

Operating contract

Main flow

Preflight

1. Resolve the target root with git rev-parse --show-toplevel when inside a repo; otherwise use the current working directory.
2. Ensure the shared .session/ workspace through session-orchestrate/scripts/session_workspace.py; never create a parallel checkpoint directory.
3. If the current session had material, preventable friction, run session-introspection first and keep only its highest-value finding/control for the checkpoint.
4. Prefer the installed save-session shell command. If it is unavailable, reproduce its behavior with Bash only.
5. Call get_goal. If the current task has an unfinished goal that should continue in a fresh task, pass its exact objective as SAVE_SESSION_GOAL_OBJECTIVE with SAVE_SESSION_GOAL_RESUME_POLICY=ensure-active. Do not infer a goal from working_on prose, and do not checkpoint completed goals as work to reactivate. A blocked goal may use ensure-active only when it should resume after its external gate changes. Use reference-only when the objective is useful history but must not be recreated.
6. If the current session has any nuance, pass curated context via environment variables before invoking the command. Prefer explicit handoff content over inferred filesystem recency:
   • SAVE_SESSION_ROOT — explicit project root when an unrelated ancestor (for example $HOME) is also a Git worktree
   • SAVE_SESSION_GOAL_OBJECTIVE — exact objective returned by get_goal; never a shortened paraphrase
   • SAVE_SESSION_GOAL_RESUME_POLICY — ensure-active or reference-only
   • SAVE_SESSION_RESUME_WINDOW_HOURS — automatic-resume eligibility window, 1–168 hours; defaults to 24
   • SAVE_SESSION_HANDOFF_FOCUS
   • SAVE_SESSION_WORKING_ON
   • SAVE_SESSION_NEXT_ACTION
   • SAVE_SESSION_BLOCKERS
   • SAVE_SESSION_LEARNINGS
   • SAVE_SESSION_VERIFICATION_STATE
   • SAVE_SESSION_INTROSPECTION
   • SAVE_SESSION_RELEVANT_FILES
   • SAVE_SESSION_RELEVANT_COMMITS
   • SAVE_SESSION_AVOID
   • SAVE_SESSION_USEFUL_COMMANDS
   • SAVE_SESSION_ROUTE_ID
   • SAVE_SESSION_ROUTE_OWNER
   • SAVE_SESSION_FIRST_COMMAND
   • SAVE_SESSION_STOP_CONDITION
   • SAVE_SESSION_ALLOWED_BEFORE_FIRST_COMMAND
   • SAVE_SESSION_AVOID_BEFORE_FIRST_COMMAND
7. Raw git status is intentionally omitted by default. Set SAVE_SESSION_INCLUDE_GIT_STATUS=1 only when the raw status itself is useful to the next agent.

Do

save-session

Optional richer invocation:

SAVE_SESSION_HANDOFF_FOCUS="Resume from the last validated result, not from broad repo discovery." \
SAVE_SESSION_GOAL_OBJECTIVE="Complete the exact accepted objective without shrinking its exit gate." \
SAVE_SESSION_GOAL_RESUME_POLICY="ensure-active" \
SAVE_SESSION_WORKING_ON="Tightening save/resume checkpoint quality so next-session context stays small and actionable." \
SAVE_SESSION_NEXT_ACTION="Run the save-session validation script, regenerate CURRENT.md once, inspect it for noise, then commit the skill change in ~/.codex." \
SAVE_SESSION_BLOCKERS="None." \
SAVE_SESSION_LEARNINGS="Recent-file inference polluted checkpoints with cache/db artifacts; require curated relevant files instead." \
SAVE_SESSION_VERIFICATION_STATE="Wrapper patched; validation still pending." \
SAVE_SESSION_INTROSPECTION="session-introspection: prevent future noisy checkpoints by preferring curated handoff fields over filesystem recency." \
SAVE_SESSION_RELEVANT_FILES="$(printf '%s\n' "$HOME/.codex/skills/save-session/SKILL.md" "$HOME/.local/bin/save-session")" \
SAVE_SESSION_AVOID="Do not include cache files, generated DB files, broad untracked status, or historical discussion not needed for the next action." \
save-session

What To Capture

Capture only context that changes the next agent's first 5 minutes:

• the active objective and why it matters now
• the exact unfinished Codex goal and whether resume should ensure it is active
• a bounded resume window appropriate for tactical state; old checkpoints must be reviewed, not auto-restored
• the first concrete next action
• a known route owner, first command, and stop condition when the next lane is already known
• validation already run and validation still pending
• the session-introspection finding/control when this session exposed preventable friction
• unresolved blockers or decisions
• files, docs, commits, branches, or commands that are directly relevant
• caveats that prevent the next agent from following stale or tempting-but-wrong context

What To Omit

Omit anything the next agent can cheaply retrieve or should not trust without rechecking:

• generated caches, databases, build outputs, logs, and broad recent-file lists
• raw git status for large dirty repos unless status ordering itself matters
• long prose history, old alternatives, and completed debate
• durable doctrine that belongs in context graph, workflow docs, AGENTS.md, project docs, or memory
• speculative conclusions that were not verified

Closeout

1. If the session produced an accepted durable repo decision and the repo has a context-graph recorder such as ./script/project_context.sh record-decision, record that decision before saving the tactical checkpoint.
2. If the session touched multiple docs, routing surfaces, AGENTS files, workflows, skills, or durable decisions, run a cheap read-only subagent or equivalent scan for stale prose, stale routing, duplicate routing, and owner conflicts. The scan returns only candidate file/line/reason; the main agent decides edits. Skip this for small single-surface handoffs.
3. Remove or demote confirmed stale/duplicate routing before writing the checkpoint, then run the narrowest available validator for touched surfaces.
4. Include the durable decision key in SAVE_SESSION_LEARNINGS, SAVE_SESSION_NEXT_ACTION, or SAVE_SESSION_ROUTE_OWNER when it changes what the next agent should retrieve first.
5. Confirm the command reported the output path.
6. Inspect ## codex_goal: its objective must exactly match get_goal, or be explicitly empty/reference-only. Never silently shorten completion criteria.
7. If useful, mention the path in the final response so the next session can resume-session.
8. Do not treat the checkpoint as durable project memory; it is tactical handoff only.

Friction introspection

Hard rules

1. Write tactically, not permanently. This skill is for handoff context, not long-term doctrine.
2. Prefer repo root. Store the checkpoint at the git root when one exists so future sessions find it predictably.
3. Stay concise. The checkpoint should guide the next move, not recreate the entire session.
4. Use Bash for the actual write. The deterministic command is the source of truth for file creation.
5. Prefer curated context over inferred context. Auto-detected files and raw status are fallback diagnostics, not handoff content.
6. Known routes need first commands. When the next lane is known, save the route owner and first command instead of prose like "run normal repo retrieval." If SAVE_SESSION_FIRST_COMMAND is set and no explicit SAVE_SESSION_NEXT_ACTION is supplied, the wrapper must derive next_action from that first command instead of using generic retrieval prose. If SAVE_SESSION_HANDOFF_FOCUS is also absent, derive handoff_focus from the first command so resume starts narrow.
7. Prevent recurrence before handoff. When the session found avoidable friction, run session-introspection before writing the checkpoint and keep only its compact control decision.
8. Do not delegate checkpoint judgment. Subagents may only scan for stale or duplicate context; the main agent records durable decisions, edits owner surfaces, validates, and writes the checkpoint.
9. Goals are explicit and exact. Save only a goal returned by get_goal or explicitly supplied by the operator. Never derive, summarize, or expand it.
10. Completion state matters. Use ensure-active for incomplete work intentionally eligible for continuation, including a blocked goal that may resume after its external gate changes. Use reference-only for completed, abandoned, or superseded work.
11. Resumability expires. ensure-active is only a candidate within SAVE_SESSION_RESUME_WINDOW_HOURS; it never makes tactical state permanently authoritative.

Cross-references

• scripts/validate.sh — audit wrapper for this skill
• scripts/save-session — committed source for the deterministic write implementation
• scripts/test_save_session.sh — deterministic goal-checkpoint tests
• ~/.local/bin/save-session — deterministic write implementation
• sibling skill: resume-session

Why this skill exists

This prevents a common failure mode at session boundaries: the next agent can see the files and git diff, but not the exact tactical intent or the first best next step. A compact repo-local checkpoint is cheaper and more reliable than reconstructing that context from scratch.