inject（inject 搜索）

DEPRECATED (removal target: v3.0.0) — Use ao lookup --query "topic" for on-demand learnings retrieval, or see .agents/AGENTS.md for knowledge navigation. This skill and the ao inject CLI command still work but are no longer called from hooks or other skills.

Inject Skill

On-demand knowledge retrieval. Not run automatically at startup (since ag-8km).

Inject relevant prior knowledge into the current session.

How It Works

In the default manual startup mode, MEMORY.md is auto-loaded by Claude Code and no startup injection occurs. Use /inject or ao inject for on-demand retrieval when you need deeper context.

In lean or legacy startup modes (set via AGENTOPS_STARTUP_CONTEXT_MODE), the SessionStart hook runs:

# lean mode (MEMORY.md fresh): 400 tokens
ao inject --apply-decay --format markdown --max-tokens 400 \
  [--bead <bead-id>] [--predecessor <handoff-path>]

# legacy mode: 800 tokens
ao inject --apply-decay --format markdown --max-tokens 800 \
  [--bead <bead-id>] [--predecessor <handoff-path>]

This searches for relevant knowledge and injects it into context.

Work-Scoped Injection

When --bead is provided (via HOOK_BEAD env var from Gas Town):

• Learnings tagged with the same bead ID get a 1.5x score boost
• Learnings matching bead labels get a 1.2x boost
• Untagged learnings still appear but ranked lower

Predecessor Context

When --predecessor is provided (path to a handoff file):

• Extracts structured context: progress, blockers, next steps
• Injected as "Predecessor Context" section before learnings
• Supports explicit handoffs, auto-handoffs, and pre-compact snapshots

Manual Execution

Given /inject [topic]:

Step 1: Search for Relevant Knowledge

With ao CLI:

ao inject --context "<topic>" --format markdown --max-tokens 1000

Without ao CLI, search manually:

# Global operating memory
sed -n '1,120p' ~/.agents/MEMORY.md 2>/dev/null

# Recent learnings
ls -lt .agents/learnings/ | head -5

# Recent patterns
ls -lt .agents/patterns/ | head -5

# Recent research
ls -lt .agents/research/ | head -5

# Global learnings (cross-repo knowledge)
ls -lt ~/.agents/learnings/ 2>/dev/null | head -5

# Global patterns (cross-repo patterns)
ls -lt ~/.agents/patterns/ 2>/dev/null | head -5

# Legacy patterns (read-only fallback, no new writes)
ls -lt ~/.claude/patterns/ 2>/dev/null | head -5

Step 2: Read Relevant Files

Use the Read tool to load the most relevant artifacts based on topic.

Step 3: Summarize for Context

Present the injected knowledge:

• Global principles or constraints that apply everywhere
• Key learnings relevant to current work
• Patterns that may apply
• Recent research on related topics

Step 4: Record Citations (Feedback Loop)

After presenting injected knowledge, record which files were injected for the feedback loop:

mkdir -p .agents/ao
# Record each injected learning file as a citation
for injected_file in <list of files that were read and presented>; do
  echo "{\"artifact_path\": \"$injected_file\", \"cited_at\": \"$(date -Iseconds)\", \"session_id\": \"$(date +%Y-%m-%d)\", \"workspace_path\": \"$PWD\"}" >> .agents/ao/citations.jsonl
done

Citation tracking enables the feedback loop: learnings that are frequently cited get confidence boosts during /post-mortem, while uncited learnings decay faster.

Knowledge Sources

Decay Model

Knowledge relevance decays over time (~17%/week). More recent learnings are weighted higher.

Key Rules

• Runs automatically - usually via hook
• Context-aware - filters by current directory/topic
• Token-budgeted - respects max-tokens limit
• Recency-weighted - newer knowledge prioritized

Examples

SessionStart Hook Invocation (lean/legacy modes only)

Hook triggers: session-start.sh runs at session start with AGENTOPS_STARTUP_CONTEXT_MODE=lean or legacy

What happens:

1. Hook calls ao inject --apply-decay --format markdown --max-tokens 400 (lean) or --max-tokens 800 (legacy)
2. CLI searches .agents/learnings/, .agents/patterns/, .agents/research/ for relevant artifacts
3. CLI applies recency-weighted decay (~17%/week) to rank results
4. CLI outputs top-ranked knowledge as markdown within token budget
5. Agent presents injected knowledge in session context

Result: Prior learnings, patterns, research automatically available at session start without manual lookup.

Note: In the default manual mode, MEMORY.md is auto-loaded by Claude Code and this hook emits only a pointer to on-demand retrieval commands (ao search, ao lookup).

Manual Context Injection

User says: /inject authentication or "recall knowledge about auth"

What happens:

1. Agent calls ao inject --context "authentication" --format markdown --max-tokens 1000
2. CLI filters artifacts by topic relevance
3. Agent reads top-ranked learnings and patterns
4. Agent summarizes injected knowledge for current work
5. Agent references artifact paths for deeper exploration

Result: Topic-specific knowledge retrieved and summarized, enabling faster context loading than full artifact reads.

Troubleshooting