[jacklau04]

Opening a coding session currently tells Claude and Codex to run /continue, even though that workflow distills memory and starts the top queued task. A direct request can therefore pick up unrelated work. Once a task is selected, continue/work also require 3–5 pages and roughly 500–1000 words of context, even when a smaller brief already contains every relevant constraint.

This change keeps project-brain retrieval and queue inspection at startup, while reserving the full resume workflow for the user's request. It replaces page/word floors with a completeness bar: objectives, acceptance criteria, decisions, constraints, current state, failed approaches, remaining work, and dated sources. Complex tasks can still carry substantial context. The Claude startup nudge uses the same boundary.

Preference curation now preserves the scope of recurring requests instead of promoting an incident, project rule, or model-specific workaround into an always-on global instruction. The new context guide describes a model-independent strategy and a representative-task evaluation method. Personal preferences, provider settings, captured data, dashboard changes, and accounting changes are not included.

Validation:

• go test -race ./internal/install ./internal/hooks ./internal/skilltest passed, including generated host instructions, startup nudge, installation/refresh, and skill protocol checks.
• go vet ./... passed; continue, work, and distill passed the skill validator; git diff --check passed.
• This validates the instruction and wiring contracts, not measured model-quality equivalence or end-to-end token savings. The documentation proposes evaluating those separately with stable acceptance criteria.

This follows OpenAI's Astra guidance on precise triggers and task-relevant reading, and Anthropic's Fable 5 guidance on revisiting prescriptive scaffolding while keeping useful memory. Model effort and verification policy remain configurable rather than being standardized across providers.