301-sf-context - SKILL.md Agent Skill

name: 301-sf-context description: "Prime task context with cached memory and focused files." argument-hint:

Canonical Paths

Before resolving any ShipFlow-owned file, load $SHIPFLOW_ROOT/skills/references/canonical-paths.md ($SHIPFLOW_ROOT defaults to $HOME/shipflow). ShipFlow tools, shared references, skill-local references/*, templates, workflow docs, and internal scripts must resolve from $SHIPFLOW_ROOT, not from the project repo where the skill is running. Project artifacts and source files still resolve from the current project root unless explicitly stated otherwise.

Chantier Tracking

Trace category: non-applicable. Process role: helper.

This skill does not write to chantier specs. If invoked inside a spec-first flow, do not modify Skill Run History; include Chantier: non applicable or Chantier: non trace in the final report when useful, with the reason and the next lifecycle command if one is obvious.

Context

Current directory: !pwd
Project name: !basename $(pwd)

Your task

Prime the session context before starting work. This avoids wasting tokens on broad file exploration.

301-sf-context answers one priming question:

What is the minimum focused context we should load before starting this known task?

Keep the boundary explicit: 301-sf-context prepares context for a known task. It does not own the final route selection, execute the lifecycle work itself, or act as a repo-status dashboard.

Route away instead of staying in 301-sf-context when the operator really needs:

skill choice or workflow routing -> 000-shipflow or 302-sf-help
actual continuation of a resolved work item -> 706-continue
cross-project git/sync reporting -> 308-sf-status

Step 1 — Check session memory

Call context_continue with the user's task as the query ($ARGUMENTS).

Read the response carefully:

Files already in memory: do NOT re-read these — use their previews
Decisions already stored: incorporate these into your plan
Budget remaining: note how much read budget is left this turn

Step 2 — Retrieve relevant files

If context_continue shows no prior memory OR the task needs more files, call context_retrieve with $ARGUMENTS as the query.

Read the ranked file list. Pick the top 3-5 most relevant files only.

Step 3 — Read only what you need

For each file you actually need to understand:

Use context_read with the query set to $ARGUMENTS (enables smart excerpting)
Skip files where the context_continue preview is already sufficient
Stop reading when budget drops below 4,000 chars

Step 4 — Report and plan

Output a concise summary:

## Ready to work on: [task]

**Files in context:**
- [file] — [one line on what's relevant]

**Key constraints/decisions:**
- [any stored decisions relevant to this task]

**Plan:**
1. [step]
2. [step]

**Budget used:** X / 18,000 chars

Then use the runtime's structured question tool when available, or a concise plain-text question:

Question: "Le contexte est prêt. On fait quoi ?"
multiSelect: false
Options:
- Proceed now (recommandé) — "Tu exécutes le plan avec ce contexte"
- Ajouter 1 fichier clé — "Tu enrichis le contexte d'un seul fichier supplémentaire"
- Affiner la cible — "Tu reformules la tâche avant d'aller plus loin"

Rules

Do NOT read files not returned by context_retrieve or context_continue
Do NOT do Glob/Grep before Steps 1 and 2
Do NOT read more than 5 files in this priming phase
If the task mentions a specific file path, add it to the read list regardless of retrieve results
If $ARGUMENTS is empty, ask the user: "What do you want to work on?" before proceeding