context-weaving

Context Weaving Skill

Unified context management combining overflow prevention, compression strategies, and filesystem offloading into a single skill. Prevents both context overflow (too many tokens from agent returns) and context rot (attention degradation in long sessions).

Core Principle

Control what returns to the Tarnished context, not what agents produce.

Agents can write unlimited detail to files. The overflow comes from what they return to the caller. This skill ensures returns are minimal (file path + 1-sentence summary).

Seven Layers of Context Management

Layer	Problem	Solution	When
Overflow Prevention	Agent returns flood lead context	Glyph Budget: file-only output	Before any Rune multi-agent command
Inter-Agent Output Compression	Ash file outputs too verbose for Runebinder	Phase 5.0 Pre-Aggregate: marker-based extraction	Before aggregation (threshold-gated)
Context Rot	Attention degrades in long contexts	Instruction anchoring, re-anchoring signals	Always (in prompts)
Compression	Session grows beyond 50+ messages	Anchored iterative summarization	During long sessions
Filesystem Offloading	Tool outputs consume 83.9% of context	Write outputs to files, read on demand	During any workflow
Compaction Recovery	Auto-compaction truncates earlier context	PreCompact checkpoint + SessionStart recovery	During long arc/arc-batch sessions
Runtime Context Monitoring	Context exhaustion during active workflows	Statusline bridge + PostToolUse warnings	Any session with monitoring enabled

Layer 1: Overflow Prevention (Glyph Budget)

When to Use

Condition	Action
Any Rune multi-agent command	Agent Teams + Glyph Budget + inscription (loaded automatically)
Custom multi-agent orchestration (3+ agents)	Agent Teams + Glyph Budget + inscription

Pre-Summon Checklist (8 Thoughts)

BEFORE summoning agents, complete the 8-step checklist: Count & Estimate → Choose Strategy → Plan Output Dir → Verify Protocol Injection → Post-Completion Validation → Revision Checkpoint → Fallback Strategies → Verification Planning.

See pre-summon-checklist.md for the full checklist with templates.

IPC Token Cost Awareness

The Glyph Budget is one half of context overflow prevention — it controls what returns to the orchestrator. The other half is IPC token cost accounting, which controls what goes into agent spawn prompts. Bloated spawn prompts displace reasoning capacity: every token of inlined content (file bodies, plan text, criteria YAML) is a token the agent cannot use for analysis. Apply the Reference Don't Inline principle — pass file paths, not file contents — to keep spawn prompts under 500 tokens. See spec-continuity.md § 8.7 for the five IPC principles.

Glyph Budget Protocol

Inject this text into EVERY agent prompt when summoning in a multi-agent workflow:

GLYPH BUDGET PROTOCOL:
- Write ALL detailed findings to: {output-directory}/{agent-name}.md
- Return to caller ONLY: the output file path + 1-sentence summary (max 50 words)
- DO NOT include full analysis, code examples, or detailed recommendations in return message

Example return:
"Findings written to {output-directory}/{agent-name}.md. Found 2 P1 issues (SQL injection, missing auth) and 4 P2 issues across 3 files."

Runebinder Pattern (Aggregation)

After all agents complete, summon the Runebinder to process raw files:

Raw files	Action
1-3 files	Read directly
4+ files	Summon Runebinder agent
10+ files	MUST summon Runebinder (never read all directly)

After aggregation, read ONLY the TOME.md file. Do NOT also read raw files.

Quality Probes (Post-Aggregation)

Probe	Check	If Failed
Circuit Breaker	ALL output files missing?	Systemic failure — abort
Inscription Validation	All expected files exist and > 100 bytes?	Report gaps
Agent Count	Does TOME.md mention all summoned agents?	Read missing agent's raw file
P1 Completeness	Are P1 findings specific (file path + line)?	Spot-check one raw file
Truthbinding Spot-Check	Verify Rune Traces in 1-2 P1 findings per agent	Compare against actual source

Token Budget Reference

Component	Tokens	Notes
CLAUDE.md + rules + MCP tools	~30k	Fixed, always present
Each agent return (without budget)	3-5k	The overflow source
Each agent return (with budget)	~100-200	File path + 50-word summary
TOME.md (aggregated)	~1k	Replaces reading all raw files
Context window	Full	Claude's per-session limit

Layer 1.5: Inter-Agent Output Compression (Pre-Aggregation)

Problem: Ash output files are verbose — 15-25% is non-finding overhead (Self-Review Log, Unverified Observations, boilerplate). When 5-8 Ashes produce 100-270 lines each, combined size exceeds 30KB, causing Runebinder stalling.

Solution: Phase 5.0 (Pre-Aggregate) extracts structured findings from Ash outputs before Runebinder ingestion. Deterministic marker-based extraction — no LLM call.

Trigger: Combined Ash output size >= review.pre_aggregate.threshold_bytes (default 25KB)

Mechanism:

1. Measure combined Ash output file size
2. If under threshold → skip (fast path, exact existing behavior)
3. If over threshold → for each Ash file:
   • Extract RUNE:FINDING marker blocks (full fidelity for P1/P2)
   • Extract ## Reviewer Assumptions and ## Summary
   • Apply priority-based trace truncation (P3: truncate to 3 lines; N: 1-line summary)
   • Discard: Self-Review Log, Unverified Observations, methodology prose
4. Write condensed files to {output_dir}/condensed/
5. Runebinder reads from condensed/ instead of raw output directory

Expected reduction: 40-60% byte savings (non-finding sections + trace truncation)

Config:

review:
  pre_aggregate:
    enabled: true               # Master toggle (default: true)
    threshold_bytes: 25000      # Auto-trigger threshold (default: 25KB)
    preserve_priorities: [P1, P2]  # Full preservation priorities
    truncate_trace_lines: 3     # Max Rune Trace lines for P3
    nit_summary_only: true      # 1-line summaries for N findings

See roundtable-circle/references/pre-aggregate.md for the extraction algorithm.

Layer 2: Context Rot Prevention

The Problem

Context rot occurs when the model's attention degrades on important instructions placed in the middle of long contexts (Lost-in-Middle effect). This is different from overflow — it happens even within token limits.

Prevention Strategies

Instruction Anchoring: Duplicate critical rules at BEGINNING and END of prompts.

# ANCHOR — TRUTHBINDING PROTOCOL
[critical rules here]

## TASK
[main instructions]

# RE-ANCHOR — TRUTHBINDING REMINDER
[repeat critical rules]

Read Ordering: Source files FIRST, agent references LAST. Keeps review criteria fresh near output generation.

Re-anchoring Signals: After every 5 files reviewed, re-check evidence rules.

Context Budget per Teammate:

• Backend review: max 30 source files
• Security review: max 20 files (all types)
• Frontend review: max 25 source files
• Docs review: max 25 markdown files

Layer 3: Compression

When to Use

Message Count	Action
< 30	No compression needed
30-49	Optional — available if needed
50-69	Trigger compression
>= 70	Force compression

Compression Strategy: Anchored Iterative Summarization

Identify truncation point → Extract key info (files, tests, decisions, errors) → Merge into existing summary → Write to tmp/scratch/session-{timestamp}.md → Continue. See context-compression.md for session summary format and quality verification dimensions.

Layer 4: Filesystem Offloading

Core Principle

Tool outputs consume up to 83.9% of context. Offload large outputs to files.

When to Offload

Output Size	Action
< 10 lines	Keep inline
10-50 lines	Consider offloading
> 50 lines	MUST offload to file

Offload Pattern

1. Write output to tmp/{workflow}/{descriptive-name}.md
2. Reference: "Output written to {path}. Key finding: {1-sentence summary}"
3. Read file only when needed for next step

Layer 5: Compaction Recovery

The Problem

Claude's auto-compaction fires during long sessions (especially /rune:arc with 43 phases), truncating earlier context. This can cause the Tarnished to lose awareness of active teams, task state, workflow phase, and arc checkpoint — leading to orphaned teams, duplicated work, or pipeline restarts.

How It Works

Two hooks form a checkpoint/recovery pair:

1. PreCompact (scripts/pre-compact-checkpoint.sh) — Fires before compaction (manual or auto). Captures a snapshot of:

   • Team config.json (active team name, members, creation time)
   • Task list state (pending/in_progress/completed counts)
   • Current workflow phase (from arc checkpoint if available)
   • Arc checkpoint path and metadata
   • Arc-batch state: current iteration, total plans, latest summary path (v1.72.0)
   • Writes to tmp/.rune-compact-checkpoint.json

2. SessionStart:compact (scripts/session-compact-recovery.sh) — Fires when session resumes after compaction. Re-injects the saved checkpoint as additionalContext so the Tarnished can resume seamlessly. Includes a correlation guard that verifies the team still exists before injection. One-time use — deletes the checkpoint file after injection to prevent stale state.

Three Ground Truth Sources

After compaction recovery, the Tarnished reconciles three sources:

Source	Contains	Priority
Team config.json	Team name, members, creation time	Authoritative for team existence
Task list	Task status, assignments, dependencies	Authoritative for work state
Arc checkpoint	Phase, artifacts, SHA-256 hashes	Authoritative for pipeline progress

Layer 5.1: Arc-Batch Inter-Iteration Summaries (v1.72.0)

When arc-batch is active, the Stop hook writes structured summary files between iterations to tmp/arc-batch/summaries/iteration-{N}.md. These summaries are captured in the pre-compact checkpoint (arc_batch_state field) and referenced in compact recovery context. Summary files are on disk (not in context), so they survive compaction and are available for Read() after recovery.

The compact recovery message includes batch iteration number, total plans, and the path to the latest summary file — giving the Tarnished immediate awareness of batch progress after compaction.

Edge case: During arc-batch, teams are created and destroyed per phase. Compaction may hit between phases when no team is active. The pre-compact hook captures batch state even without an active team, writing a minimal checkpoint. The recovery hook handles teamless checkpoints by injecting batch context directly without the team correlation guard.

Relationship to Rule #5

This layer automates what Rule #5 ("On compaction or session resume: re-read team config, task list, and inscription contract") previously required manually. The PreCompact hook captures state proactively, and the SessionStart:compact hook re-injects it — ensuring Rule #5 compliance even when compaction truncates the original context.

Layer 6: Runtime Context Monitoring (v1.78.0)

The Problem

Even with overflow prevention and compression, a session can creep toward context exhaustion during long workflows without visible feedback. By the time the model notices degraded output quality, significant context has already been lost.

Components

• Statusline Bridge — A statusline script that writes context metrics (percentage used, token counts) to a shared bridge file at /tmp/rune-ctx-{session_id}.json. Runs as a color-coded statusline with git branch and active workflow detection. The bridge file is the producer half of a producer/consumer pattern.
• Context Monitor — A PostToolUse hook (scripts/rune-context-monitor.sh) that reads the bridge file and injects agent-visible warnings when context usage crosses thresholds. Non-blocking (exits 0). Only injects when the bridge file is fresh (staleness guard: 5 minutes). The monitor is the consumer half of the pattern.
• Plan Budget — An optional session_budget frontmatter field in plan files that caps simultaneous agent spawning (max_concurrent_agents). Validated silently by strive/arc worker orchestration to prevent context saturation from large teams.

Setup

Enable in .rune/talisman.yml:

context_monitor:
  enabled: true                    # Master toggle for context monitoring
  warning_threshold: 35            # Warn when remaining% <= this (default: 35)
  critical_threshold: 25           # Critical stop when remaining% <= this (default: 25)
  stale_seconds: 60                # Bridge file max age before ignoring (default: 60)
  debounce_calls: 5                # Tool uses between repeated warnings (default: 5)
  workflows: [review, audit, work, mend, arc, devise]  # Which workflows emit warnings

Thresholds

Level	Context Remaining	Injected Message
WARNING	≤ 35%	"Context at {pct}% remaining. Consider compacting or reducing agent scope."
CRITICAL	≤ 25%	"Context CRITICAL at {pct}% remaining. Compact now or risk truncation."

Architecture Notes

• Producer/Consumer pattern: statusline writes, monitor reads via /tmp/ bridge file. No shared memory — pure filesystem.
• Session-isolated: bridge files are keyed by session_id with config_dir + owner_pid ownership fields.
• Non-blocking: all errors exit 0 — the monitor never blocks tool execution.
• Bridge file cleanup: on-session-stop.sh scans /tmp/rune-ctx-*.json and removes files matching the current session's ownership markers.

Layer 6.1: Glyph Budget Hook Enforcement (v1.102.0)

A PostToolUse:SendMessage hook (scripts/enforce-glyph-budget.sh) reinforces the Glyph Budget Protocol at the infrastructure layer. When any teammate sends a SendMessage call exceeding the word budget, the hook injects an advisory context message.

Threshold: 300 words (configurable via RUNE_GLYPH_BUDGET environment variable or talisman context_weaving.glyph_budget.word_limit).

Behavior: Advisory-only — PostToolUse cannot block. The advisory informs the orchestrator for corrective action on future messages.

Guard: Only active when Rune workflow state files exist in tmp/ (detection via explicit file path checks — not globs).

# talisman.yml schema (optional config)
context_weaving:
  glyph_budget:
    enabled: true              # Toggle enforcement
    word_limit: 300            # Max words before advisory
    enforcement: advisory      # advisory | warning | block (future)

Layer 6.2: Adaptive Context Degradation (3-Tier System, v1.102.0)

The guard-context-critical.sh PreToolUse hook now implements a 3-tier context degradation system that provides graduated responses as context approaches exhaustion:

Tier	Remaining %	Action	Block?
Caution	35–40%	Advisory: compress messages, prefer file output	No
Warning	25–35%	Workflow-specific degradation suggestions	No
Critical	≤25%	Hard DENY for TeamCreate/Agent (except Explore/Plan)	Yes

Caution Tier (40% remaining)

Injects additionalContext with general compression suggestions. No block. Applies to all TeamCreate/Agent calls.

Warning Tier (35% remaining)

Detects active workflow from state files (with session ownership check) and injects workflow-specific degradation suggestions:

Workflow	Suggestion
review/appraise/audit	Reduce team to 3-4 Ashes
work/strive	Complete current task, skip optional tasks
arc	Skip optional phases, proceed to ship
devise/plan	Skip forge enrichment
unknown	Reduce scope, prefer file-based output

Critical Tier (25% remaining, unchanged)

Hard DENY for TeamCreate and Agent tool calls. Explore/Plan agents remain exempt (read-only, minimal context cost). Escape hatches: /rune:rest, talisman disable flag, or Explore/Plan agent types.

References

• Overflow Wards — Detailed pre-summon protocol
• Context Compression — Compression strategies and triggers
• Filesystem Offloading — File-based output patterns
• Companion: rune-orchestration (architectural patterns, inscription protocol)