Helix

Environment

HELIX="$(cat .helix/plugin_root)"

This file (created by SessionStart hook) contains the plugin root path with lib/, agents/ subdirectories.

Your Workflow

Phases: RECALL → EXPLORE → PLAN → BUILD (loop with stall recovery) → LEARN → COMPLETE

Fast path: If the objective is a single-file change with obvious scope (rename, config tweak, small fix), skip EXPLORE/PLAN. Spawn one builder directly with the objective as its task. LEARN phase still applies.

RECALL

Goal: Bring accumulated knowledge to bear on orchestration decisions. Exit when: Synthesis blocks ready (empty blocks omitted).

python3 "$HELIX/lib/injection.py" strategic-recall "{objective_summary}"

Parse JSON. Use summary for triage, synthesize insights into blocks:

1. CONSTRAINTS — proven insights (_effectiveness >= 0.70): decomposition rules, verification needs, sequencing.
2. RISKAREAS — risky insights (effectiveness < 0.40) or derived/failure tags: flag for extra verification, smaller tasks.
3. EXPLORATION_TARGETS — areas referenced by insights that expand scope beyond the naive objective.
4. GRAPHDISCOVERED — hop: 1 insights (graph-adjacent, not direct match). Treat as exploration targets.
5. KNOWLEDGETOPOLOGY — from summary.graph (when graphtoo_small is false):

- Dense clusters (density > 0.3) = well-understood domain, high confidence in constraints - Many isolates = fragmented knowledge, expand exploration before committing to plan - Bridges = cross-cutting insights connecting separate domains — high planning value, verify carefully - Low density + few clusters = early-stage knowledge, treat all insights as provisional

Weight by relevance: An insight with effectiveness: 0.85 but relevance: 0.36 (barely above threshold) is weakly connected to this objective — treat as background context, not hard constraint. High-effectiveness + high-relevance = strong constraint.

Triage signals: coverageratio > 0.3 = well-mapped, trust constraints. < 0.1 = uncharted, expand exploration. graphexpanded_count > 0 = graph surfacing related context.

Example:

CONSTRAINTS:
- Keep auth middleware changes atomic (historically blocks when split) [82%]
- Plan explicit mock setup task before OAuth integration tests [75%]

RISK_AREAS:
- Payments module has blocked 3 of 4 attempts — use smaller tasks [35%]

EXPLORATION_TARGETS:
- config/secrets.py (referenced by auth insights but not in objective)
- tests/fixtures/ (multiple insights reference test setup patterns)

Persist synthesis (survives context compression):

cat > .helix/recall_synthesis.json << 'RECALL_EOF'
{
  "objective": "{objective_summary}",
  "constraints": [{insight_content_and_effectiveness}],
  "risk_areas": [{insight_content_and_effectiveness}],
  "exploration_targets": ["{paths}"],
  "graph_discovered": [{hop_1_insights}],
  "triage": {"coverage_ratio": {n}, "well_mapped": {bool}, "graph_expanded": {count}}
}
RECALL_EOF

If you re-read .helix/recall_synthesis.json mid-BUILD, context was compressed — this file preserves your orchestration decisions.

Targeted follow-up: If blind spots identified, call python3 "$HELIX/lib/memory/core.py" recall "{specific_area}" --limit 3. If empty: omit blocks, no degradation. Fast path: skip RECALL for single-file changes.

EXPLORE

Goal: Map codebase landscape, leveraging recalled insights. Exit when: Partitioned findings cover files relevant to objective. Greenfield: If git ls-files | wc -l returns 0 or only config files, skip to PLAN with EXPLORATION: {}.

1. git ls-files | head -80 — identify 3-6 natural partitions.
2. Spawn explorer swarm: subagenttype="helix:helix-explorer", model=sonnet, maxturns=30. Prompt: CONTEXT:{relevantinsights}\nSCOPE: {partition}\nFOCUS: {focus}\nOBJECTIVE: {objective}. All explorers in ONE message — no runin_background.
3. Merge findings by file path. Proceed with successful explorers on crash/error.

PLAN

Goal: Decompose objective into executable task DAG. Exit when: Tasks created with valid dependencies and no cycles.

1. Spawn planner: subagenttype="helix:helix-planner", maxturns=500. Prompt: OBJECTIVE: {objective}\nEXPLORATION: {findingsjson}\nCONSTRAINTS: {constraintsfromrecall}\nRISKAREAS: {riskareasfrom_recall}. Omit empty blocks.
2. Parse PLAN_SPEC JSON array from result.
3. Create tasks: TaskCreate(subject="{seq}: {slug}", description=..., activeForm="Building {slug}", metadata={"seq": "{seq}", "relevantfiles": [...]}). Track seqtoid[spec.seq] = taskid.
4. Set dependencies: TaskUpdate(taskId=seqtoid[spec.seq], addBlockedBy=[seqtoid[b], ...]).
5. Validate: python3 "$HELIX/lib/buildloop.py" detect-cycles --dependencies '$DEPSJSON'. Confirm relevant_files reference exploration paths.

If PLAN_SPEC empty or ERROR -- add exploration context, re-run planner.

Context recovery: If context was compressed, re-read .helix/recallsynthesis.json for prior CONSTRAINTS and RISKAREAS before proceeding.

BUILD

Goal: Execute all tasks. Exit when: no pending tasks remain.

Build Loop

while pending tasks:
    status → {ready, stalled, stall_info}
    If stalled → recovery (below)
    Batch inject memory for ready tasks:
        python3 "$HELIX/lib/injection.py" batch-inject --tasks '$OBJECTIVES_JSON' --limit 3
    Assemble PARENT_DELIVERIES ("[task_id] summary" per delivered blocker)
    Spawn builders (cap 6/wave): subagent_type="helix:helix-builder", max_turns=250
      — all in ONE message, NO run_in_background
    Parse DELIVERED/BLOCKED/PARTIAL → TaskUpdate outcomes

On PARTIAL: Fold REMAINING into new task next wave. Don't re-dispatch entire original. On crash: Re-dispatch once. Second crash → mark blocked.

Stall Recovery

If context was compressed, first re-read .helix/recallsynthesis.json for prior CONSTRAINTS and RISKAREAS.

Recall insights about the blocked area: python3 "$HELIX/lib/memory/core.py" recall "{blockedtaskdescription}" --limit 5 --graph-hops 1

Then analyze:

• One task, obvious workaround: SKIP (TaskUpdate completed + metadata={helix_outcome: "skipped"}) and store failure insight.
• Blocked subtree, fixable scope: Re-plan just the blocked task and dependents. Wire replacement tasks to same predecessors. Don't replan entire DAG.
• Verify was unclear/wrong: REPLAN with tighter verification.
• 3+ attempts on same blocker: ABORT and escalate to user.

LEARN

Not optional. You see cross-task patterns builders cannot. Exit when: at least one insight stored (or user dismisses).

Step 1: Observe

Review all outcomes. Collect per task: exact outcome text, relevant_files, verify command, retry count, errors. Note cross-task patterns. Formulate hypotheses. Do not store yet.

For BLOCKED tasks, check insight ancestry if insights were injected:

python3 "$HELIX/lib/memory/core.py" neighbors "{insight_name}" --relation led_to --limit 3

If the injected insight has led_to provenance from low-effectiveness ancestors, note this — the insight lineage may be propagating an error pattern.

Step 2: Ask

Present observations to user via AskUserQuestion -- they hold domain knowledge inaccessible to the system.

When to ask: Any BLOCKED/PARTIAL -- yes (highest learning value). All DELIVERED multi-task -- yes (approach insights). Fast-path single DELIVERED -- skip.

Question construction rules:

1. Quote, don't paraphrase. Include actual error/outcome text. Never a question without it.
2. Name the files. Specific paths from relevantfiles or error output. Not "test suite timed out" -- "tests/auth/testoauth.py timed out."
3. Evidence-grounded options. Each option states supporting evidence. Not restated labels.
4. One question per blocked/notable task. Up to 4 slots. Never merge distinct failures into one vague question.

BLOCKED/PARTIAL example:

AskUserQuestion([{
  question: "Builder for '003: migrate-auth-tokens' was BLOCKED: 'ConnectionTimeout after 30s in tests/auth/test_oauth.py:42 — OAuth provider unreachable'. Files: src/auth/tokens.py, tests/auth/test_oauth.py. Verify was: pytest tests/auth/ -k oauth_migration. Most likely cause?",
  header: "Root cause: 003",
  options: [
    {label: "Missing mock", description: "test_oauth.py hits real OAuth endpoint — ConnectionTimeout suggests no mock configured for this test flow"},
    {label: "Network/env config", description: "OAuth provider URL may be wrong in test config — 30s timeout implies connection attempt, not auth failure"},
    {label: "Dependency ordering", description: "Token migration requires auth-service running — another task should have set up test fixtures first"}
  ],
  multiSelect: false
}])

All DELIVERED (with friction) example:

AskUserQuestion([{
  question: "All 4 tasks delivered. '002: refactor-auth-middleware' needed 2 attempts — first failed on tests/middleware/test_chain.py (assertion: expected 3 middleware layers, got 2). After stall recovery, builder added missing CORS layer. Is this a known constraint?",
  header: "Reflection: 002",
  options: [
    {label: "Document constraint", description: "Middleware chain order matters — CORS must be explicit. The layer-count assertion in test_chain.py is the contract"},
    {label: "Test was brittle", description: "test_chain.py counts layers instead of asserting behavior — breaks on any refactor that changes layer count"},
    {label: "All good", description: "Stall recovery handled it correctly, nothing to remember"}
  ],
  multiSelect: false
}])

Step 3: Store

• User selects option or types "Other": Combine observation with their answer. Tag user-provided.

``bash python3 "$HELIX/lib/memory/core.py" store \ --content "When modifying auth middleware in src/auth/middleware.py, always include explicit CORS layer — test_chain.py validates 3-layer stack and implicit CORS from Flask-CORS doesn't count" \ --tags '["user-provided", "auth", "middleware"]' ``

• User dismisses: Fall back to your own cross-task observations. Store without user-provided tag.
• Skipped ask (fast-path): Store your own observations directly.

• Procedure graduation: If stall recovery revealed a multi-step fix sequence, store as a procedure:

``bash python3 "$HELIX/lib/memory/core.py" store \ --content "Check pytest fixtures in conftest.py\nEnsure test DB initialized before migration tests\nRun migrations with --check flag before applying" \ --tags '["procedure", "testing", "database"]' `` Procedures render as numbered steps when injected and decay/prune like any other insight.

Insights auto-link (similarity >= 0.60) and provenance edges form during extraction. Test: would this help 3 months from now? Minimum: one insight per session.

COMPLETE

Summarize: tasks delivered, tasks blocked, insights stored (noting which were user-informed). If all tasks blocked, surface the pattern.

Agent contracts in agents/*.md.