SageOx Summary#

You are an agent that generates a cross-team summary of the last 24 hours of SageOx distilled activity. You enumerate each team's recent distilled entries via ox distill history, inline their content into a prompt for claude -p, and return a structured Slack-formatted summary.

This skill pairs with the sageox-distill skill — distill writes the source material, this skill synthesizes it.

ox version requirement: this skill uses ox distill history list and ox distill history show, which landed in PR #507. If ox distill history --help returns "unknown command" or similar, update ox (see § 3 below) before continuing.

Prerequisites#

Before doing anything else, verify the user's environment. Run every check in order. If any required check fails, explain precisely what's missing and stop. Do not proceed until the user has fixed it.

1. Required binaries#

ox, claude, and jq are declared in the front matter's requires.bins, so OpenClaw checks them before running the skill. claude (npm) and jq (brew) have declarative installs in the front matter; ox does not. If OpenClaw reports a missing bin, surface its message to the user and stop — except for ox, which has the interactive install flow in § 3 below.

claude -p will use whatever credentials claude already has — either an OAuth session from claude login (Pro/Max subscription) or ANTHROPIC_API_KEY exported in the shell that launched OpenClaw. The skill no longer accepts a per-skill apiKey.

2. Path validation rules#

Several steps below read a path from a JSON state file. Before interpolating any such value into a shell command, the agent must validate it against these rules:

1. Absolute path required. Must start with / or ~. Reject relative paths and bare names.
2. No .. segments. Reject anything containing ...
3. No shell metacharacters. Reject anything containing any of these characters: ; $ ` | & < > ( ) { } * ? [ ] ! \ newline.

On any validation failure: print a clear error to the user explaining which rule failed and ask them to provide a different path. Do not attempt to "fix up" or sanitize the input — reject and re-prompt.

Treat values read from ~/.openclaw/memory/*.json files as untrusted even though this skill writes them: the user (or a process running as the user) may have edited the file by hand or by another tool between runs. Re-validate every read.

3. Installing ox#

The ox CLI install state is recorded in ~/.openclaw/memory/sageox-ox-install.json. On every run of this skill, invoke the bundled readiness gate:

bash

bash scripts/update-ox.sh

Contract:

• Stdout: nothing on success
• Stderr: on any failure, a two-line message — an error: line describing what's wrong, followed by a fix: line with the remediation. Surface both verbatim to the user.
• Exit: 0 ox is pinned, installed, and reports the expected version (continue to § 4); 2 ox is not usable — one of: state file missing, binary missing at $HOME/.local/bin/ox, ox on PATH resolves to a different binary, binary fails to run, or binary reports a version other than the one recorded in sageox-ox-install.json. On exit 2, STOP, read references/INSTALL.md, follow the install flow, then re-run this script to confirm.

There is no per-run auto-update. The curl install pins a specific ox release by tag and sha256; users pick up newer releases by re-running clawhub install for this skill after a new skill version publishes. The user can say "reinstall ox" at any time to re-enter the flow in references/INSTALL.md.

Do not install ox via Homebrew or any package manager (e.g. brew install sageox/tap/ox, apt, dnf, pacman). The tap exists for general use but is not supported inside OpenClaw skills — only the pinned-release curl flow is.

4. Authentication#

1. ox status — confirm ox is authenticated. If not, tell the user to run ox login and try again.

2. Smoke-test claude -p: bash

   claude -p "say hi" --model claude-sonnet-4-6
   
   

   If it fails with an auth error, claude has no usable credentials. Tell the user to either run claude login (Pro/Max OAuth) or export ANTHROPIC_API_KEY in the shell that launches OpenClaw, then re-run the skill. The skill cannot inject the key itself.

Configuration#

The skill uses two pieces of state:

1. Repo manifest — ~/.openclaw/memory/sageox-distill-repos.json (shared with the sageox-distill skill). Format: json

   {
     "repos": [
       { "path": "/home/user/repos/my-project", "team_id": "my-team" }
     ]
   }
   
   

   The team_id values are what drive the summary — they're passed directly to ox distill history list --team. The path entries are not used by this skill, but stay in the manifest so sageox-distill keeps working.

2. Summary state — ~/.openclaw/memory/sageox-summary-state.json. Tracks which distilled daily entry ids have already been included in a prior summary run so the skill never re-summarizes the same content. Shape: {updated_at, teams: {<team_id>: {included_ids: [<id>, ...]}}}. Missing file → empty state (first run). Treat contents as untrusted: every id must pass ^[0-9]{4}-[0-9]{2}-[0-9]{2}-[0-9a-f-]+$ or be silently dropped. Both bundled scripts handle this already — the rule is stated here for anyone reading the state file directly.

   Pre-0.2.0 versions of this skill stored entries under included_files (with a .md suffix). That field is ignored by the current scripts; users upgrading from 0.1.x will re-summarize the most recent window once and the state file will converge to the new schema on the next successful run.

If the manifest does not exist, tell the user to run the sageox-distill skill first to set up the repos, or ask for paths directly and populate it.

The skill no longer needs a SageOx endpoint file or any team-context directory layout knowledge — ox distill history resolves paths internally. Earlier versions read ~/.openclaw/memory/sageox-endpoint.txt; that file is ignored now and can be left alone or removed.

Summary Pipeline#

When the user asks for a summary, run the steps in order. Steps 2 and 5 delegate their mechanics to scripts/select-new-entries.sh and scripts/update-state.sh (invoke via bash scripts/<name>).

Step 1: Load the Manifest and Summary State#

1. Read ~/.openclaw/memory/sageox-distill-repos.json with jq. Re-validate every path entry against the Path validation rules in Prerequisites § 2 before using it — the manifest is user-writable and may have been hand-edited between runs. (The summary pipeline itself does not dereference the paths, but if the manifest looks corrupt we stop rather than guess at intent.)

2. Read ~/.openclaw/memory/sageox-summary-state.json with jq if it exists. If it is missing, proceed silently as if the teams map were empty — first runs are normal. If it exists but is malformed or unreadable, proceed as if empty AND emit exactly one warning to stderr: text

   warning: sageox-summary-state.json was unreadable, starting from empty state
   
   

   Never route this warning to stdout — stdout is the final summary (Step 6), and mixing decorative output into it breaks downstream consumers. This file is rewritten at the end of every successful run (see Step 5).

3. Collect the unique team_id values from the manifest. These drive every subsequent ox distill history call.

Step 2: Select New Entries per Team#

The window is the last 24h, resolved server-side by ox distill history list against entry created_at timestamps. ox auto-expands the window around the UTC day boundary so runs don't miss yesterday's late entries — this skill passes 24h verbatim and lets ox decide how wide to actually look. The state file then subtracts anything already summarized, so re-runs within the window are idempotent. First run (no state file) = every candidate is new, not an error.

For each unique team_id, invoke scripts/select-new-entries.sh. It runs ox distill history list --team <tid> --since 24h --layer daily --format json, extracts the entry ids, and subtracts the team's included_ids set from the state file. Contract:

• Usage: select-new-entries.sh <team_id> <since> <state_file>
• Stdout: one entry id per line, sorted; empty if nothing new
• Stderr: one-line warnings (malformed state, failed ox call)
• Exit: 0 success (empty is not a failure; a failed ox call is treated as "no new entries" so the remaining teams still summarize), 2 usage, 3 internal ( jq or ox missing)

bash

STATE_FILE=~/.openclaw/memory/sageox-summary-state.json
NEW_IDS="$(bash scripts/select-new-entries.sh \
  "$TEAM_ID" 24h "$STATE_FILE")"

Then:

1. If NEW_IDS is empty for a team, skip it this run — log <team_id>: no new entries since last summary to stderr and continue with the remaining teams.
2. If every team ends up with zero new entries, stop the pipeline before invoking Claude. Print one line to stdout — No new distilled content since last summary. — and exit. Do not modify the state file; Step 5's prune on the next successful run will collect any stale entries.

Step 3: Fetch Entry Content and Build the Prompt#

For each team that survived Step 2, fetch the content of every new entry in a single ox distill history show call. show accepts multiple ids and emits each entry's markdown separated by its own <!-- entry: <id> --> header:

bash

# shellcheck disable=SC2086
TEAM_BLOCK="$(ox distill history show \
  --team "$TEAM_ID" \
  --format content \
  $NEW_IDS)"

(Intentionally unquoted expansion: NEW_IDS is a sorted set of ids, each already matched against ^[0-9]{4}-[0-9]{2}-[0-9]{2}-[0-9a-f-]+$ by Step 2, so word-splitting is safe and desired here.)

If the show call fails for a team, surface the stderr and stop the whole run — partial summaries misrepresent the team's activity, and Step 5 stays untouched so the next run retries the same set.

Then read the template from the skill's assets directory: ./assets/SUMMARIZE.md (relative to this SKILL.md file). The file path on disk depends on where OpenClaw loaded the skill from — typically one of:

• ~/.openclaw/skills/sageox-summary/assets/SUMMARIZE.md
• ./skills/sageox-summary/assets/SUMMARIZE.md (workspace skill)

Substitute template placeholders:

• {{ENTRIES}} — one section per team that has a non-empty NEW_IDS set, in this format: text

  ### Team "<team_id>"
  
  <TEAM_BLOCK for that team>
  
  

  Teams with zero new entries were already dropped in Step 2 and must not appear in {{ENTRIES}}. Do not re-number, re-wrap, or otherwise mutate the markdown inside TEAM_BLOCK — it is authored by ox distill and preserves the citation anchors Claude references in the final summary.

• {{MULTI_TEAM_RULES}} — if two or more teams survived Step 2 with non-empty entry sets, replace with: text

• Organize the summary by team, using each team ID as a section header

• Attribute insights to the correct team

Otherwise, replace with an empty string.

### Step 4: Run Claude

Invoke `claude -p` with the substituted prompt. The prompt already
contains every entry's full text, so Claude does not need filesystem
access — do **not** pass `--add-dir` and do **not** grant read tools.

- `--model claude-sonnet-4-6`
- No `--add-dir` (content is inline)
- No `--allowedTools` (prompt is self-contained)
- The substituted prompt passed via stdin

`claude -p` will use whatever credentials `claude` already has — either
an OAuth session from `claude login` or `ANTHROPIC_API_KEY` from the
shell that launched OpenClaw (see Prerequisites § 1). Wrap the
invocation in `timeout 600` (10 minutes) — this matches the
timeout used by `pkg/sessionsummary/claude.go` in the `ox` repo for
comparable Claude synthesis work and gives the model enough headroom for
cross-team summaries that inline many daily entries:

bash

```bash
timeout 600 claude -p \
  --model claude-sonnet-4-6 <<< "$PROMPT"

If the invocation fails (non-zero exit, timeout exit 124, network error), surface the error to the user and do not proceed to Step 5 — leaving the state file untouched lets the next run retry exactly the same candidate set.

Step 5: Update Summary State#

Only run this step if Claude exited successfully. On any failure (non-zero exit, timeout, network error), skip it entirely — leaving the state file untouched lets the next run retry the same candidate set.

For each team that had non-empty NEW_IDS in Step 2, pipe that team's ids into scripts/update-state.sh. The script merges them into the team's included_ids, prunes entries whose date prefix is strictly older than yesterday UTC (the candidate window, after ox's auto-expansion, spans at most today + yesterday), and writes the result atomically via a sibling temp file + mv -f. See the script header for full details; the short form:

• Usage: update-state.sh <state_file> <team_id>
• Stdin: one entry id per line (regex-filtered on read)
• Stdout: nothing on success
• Stderr: one-line warnings (e.g. malformed prior state)
• Exit: 0 success, 2 usage, 3 internal

bash

printf '%s\n' "$NEW_IDS" \
  | bash scripts/update-state.sh "$STATE_FILE" "$TEAM_ID"

Teams skipped in Step 2 are not invoked here — their prior state passes through unchanged because update-state.sh only rewrites the team_id it was given. Teams with no prior entry AND no new ids are correctly never added to the state file.

Step 6: Return the Summary#

Return Claude's stdout to the user directly. It is already formatted for Slack mrkdwn. Do not reformat or annotate — just show it.

Output#

The primary output is the Claude-generated summary. Prefix it with a brief one-line header showing:

• How many teams were summarized
• Any teams that were skipped (and why — typically "no new entries since last summary" from Step 2)

Keep any preamble or postamble minimal. The summary itself is the value.

If Step 2 short-circuited because every team had zero new entries, the only output is the single line No new distilled content since last summary. — do not invoke Claude and do not print a header.