feat(skills): DS-12 and DS-13 — config override application, batch auto-detection, and context preservation warning

[levxn]

Summary

Implements DS-12, DS-13, and config override application for default skills. Config overrides set in agent definitions (e.g. checkpoint_every_n: 10, warn_at_usage_ratio: 0.4) were previously parsed and logged but silently discarded — they now take effect at runtime.

• Config overrides: {{placeholder}} substitution added to SKILL.md bodies; overrides from default_skills agent config replace default values before the protocol is injected into the system prompt
• DS-12: hive.batch-ledger auto-detects batch scenarios from goal/input text at node startup and prepends a ledger-init nudge to the system prompt; configurable via auto_detect_batch override
• DS-13: hive.context-preservation monitors token usage during execution and injects a one-time preservation warning when usage crosses warn_at_usage_ratio (default 0.45), before the framework's 0.6 prune threshold fires

---

What Changed

core/framework/skills/_default_skills/quality-monitor/SKILL.md

• Replaced hardcoded 5 with {{assessment_interval}} placeholder

core/framework/skills/_default_skills/error-recovery/SKILL.md

• Replaced hardcoded 3 with {{max_retries_per_tool}} placeholder

core/framework/skills/_default_skills/context-preservation/SKILL.md

• Added a line: You will receive an alert when context reaches {{warn_at_usage_ratio_pct}}% — preserve immediately.

core/framework/skills/defaults.py

• Added _SKILL_DEFAULTS dict with per-skill default values for all placeholders
• Added _apply_overrides() — merges defaults with agent-supplied overrides and substitutes {{key}} in skill body text; handles warn_at_usage_ratio float to percent conversion
• build_protocols_prompt() now calls _apply_overrides() per skill before appending to the combined prompt
• Added is_batch_scenario(text) — keyword heuristic for DS-12 batch detection (11 keywords: "list of", "for each", "records", "items", etc.)
• Added batch_init_nudge property — returns nudge text when hive.batch-ledger is active and auto_detect_batch is not False
• Added context_warn_ratio property — returns warn_at_usage_ratio override or 0.45 default; returns None when hive.context-preservation is disabled

core/framework/skills/manager.py

• SkillsManager stores a reference to DefaultSkillManager after _do_load()
• Exposes batch_init_nudge and context_warn_ratio as passthrough properties

core/framework/runtime/agent_runtime.py

• Reads context_warn_ratio and batch_init_nudge from SkillsManager after load; stores as instance attributes and passes to ExecutionStream

core/framework/runtime/execution_stream.py

• Accepts and threads context_warn_ratio and batch_init_nudge through to GraphExecutor

core/framework/graph/executor.py

• Accepts and threads both new params through to NodeContext creation

core/framework/graph/node.py

• Added default_skill_batch_nudge: str | None and default_skill_warn_ratio: float | None to NodeContext

core/framework/graph/event_loop_node.py

• DS-12: after protocols are injected into the system prompt, checks ctx.default_skill_batch_nudge; runs is_batch_scenario() against goal_context + input_data; appends nudge if batch is detected
• DS-13: initializes _context_warn_sent = False per execution; after each successful LLM turn, checks usage_ratio() >= ctx.default_skill_warn_ratio; injects a user-facing CONTEXT ALERT message once and sets the flag

core/tests/test_default_skills.py

• 19 new tests across three new classes: TestConfigOverrideSubstitution, TestBatchAutoDetection, TestContextWarnRatio

---

Why

Agents configured with assessment_interval: 10 or warn_at_usage_ratio: 0.4 expect those values to change agent behavior — previously they were no-ops. Batch auto-detection removes the manual overhead of telling the agent to initialize a ledger every time. Context preservation warnings give the agent advance notice at 45% usage rather than waiting until the framework silently prunes at 60%.

---

How to Test

Run the updated test file:

uv run pytest core/tests/test_default_skills.py -v

All 44 tests pass. Full suite regression:

uv run pytest core/tests/ -q

1002 passed, 0 failures.

---

Out of Scope

checkpoint_every_n is parsed and stored as an override but has no runtime hook yet — it is included in _SKILL_DEFAULTS so it will substitute cleanly into any future SKILL.md placeholder that references it. Iteration and node-completion hook infrastructure (DS-9, DS-10) are tracked separately.

---

solves issue #6365

[levxn]

@TimothyZhang7 let me know if it works!