phrazzld/orient

/orient

Start a session from repo truth, not memory. Read-only. Fast by default.

Contract

Produce a short orientation report that prevents a wrong first move. Do not deliver, groom, debrief, assess readiness, reflect, mine transcripts, or mutate state.

Default budget: under 2 minutes, under 12 report lines. Use --deep only when the user asks for more archaeology.

Sources

Read the smallest live set that explains the current workspace:

• scoped AGENTS.md and repo AGENTS.md
• project.md, then README.md only if project focus is still unclear
• git status --short --branch --untracked-files=all
• current branch and recent commits
• active backlog.d/*.md count and titles
• newest relevant backlog.d/_done/*.md items or closure trailers
• .harness-kit/agents.yaml or roster probe output when next action depends on delegation
• .harness-kit/agent-readiness.yaml only to name the profile state, not to score the repo

Optional when requested or obviously relevant: work ledger, delegation receipt summary, skill invocation analytics, review-score trends, or docs site state.

Report Shape

**Orientation**
- Repo / branch:
- Workspace state:
- Current focus:
- Backlog signal:
- Recent closure signal:
- Roster state:
- Applicable constraints:
- Blockers / gaps:
- Likely next skill:
- Residual uncertainty:

Keep each line evidence-backed. Include exact paths or commands for claims that would change the next action.

Routing Judgment

| Signal | Recommend | |---|---| | Active shaped ticket and clean branch | /deliver <ticket> | | Dirty existing branch with intended changes | /deliver --polish-only <branch> | | Empty or stale backlog | /groom | | User asks what happened or why it matters | /debrief | | Failure, broken gate, or unclear root cause | /diagnose | | Finished work needing closeout | /ship or /reflect by context | | Readiness/profile question | /agent-readiness | | Skill/harness primitive change | /harness-engineering |

If the next action is unclear, say what evidence is missing instead of choosing a workflow from vibes.

Gotchas

• Do not turn orientation into a mandatory ceremony. If the user already gave a precise command and the workspace state is obvious, keep it tiny.
• Do not summarize all history. Name the latest useful closure signal and stop.
• Do not store session memory. Durable state belongs in backlog, receipts, traces, commits, or explicit profile files.
• Do not call provider lanes unless scope is broad, stale, or contested.
• Do not label the repo "ready" or "validated" from orientation. Route that to the owning skill.
• Do not replace scoped AGENTS.md; respect it as the governing instruction source.

Verification

When editing this skill in the Harness Kit source repo, validate trigger/catalog shape and eval structure with:

cargo run --locked -p harness-kit-checks -- check-frontmatter --repo .
cargo run --locked -p harness-kit-checks -- check-skill-evals --repo .

Semantic acceptance is the grader:

cargo run --locked -p harness-kit-checks -- eval-grader orient-session-start <candidate-output>

When invoking /orient in another repo, these source-repo commands are not required. Acceptance is the report's cited live evidence and useful next-skill recommendation.