pipeline-orchestration

name	pipeline-orchestration
description	Cross-domain loop orchestration for the chip design pipeline. Provides the fix_request protocol, iteration-cap logic, escalation templates, and dispatch patterns for routing verification/formal failures to the RTL orchestrator and back. Use when driving the closed-loop verification↔RTL feedback cycle.
version	1.0.0
author	chuanseng-ng
license	MIT
allowed-tools	Read, Write, Bash

Skill: Pipeline Orchestration

Invocation

• If invoked by a user presenting a pipeline loop task: immediately spawn the digital-chip-design-agents:pipeline-orchestrator agent and pass the full user request and any available context. Do not execute stages directly.
• If invoked inside another orchestrator: read design_state.json, summarise open fix_requests[], and return — do not spawn subagents (anti-recursion rule).

Purpose

This skill provides the closed-loop verification↔RTL feedback protocol. When a DUT bug is found during simulation or formal verification, it must be communicated to the RTL orchestrator in a machine-actionable way and the pipeline must iterate until the bug is fixed or the iteration limit is reached.

The protocol has three participants:

Participant	Role
verification-orchestrator / formal-orchestrator	Detects the bug; writes a fix_request entry to design_state.fix_requests[] with status=open; terminates with decision=escalate.
rtl-design-orchestrator	Reads the open fix_request; sets status=claimed; fixes the RTL; sets status=fixed with rtl_response; terminates.
pipeline-orchestrator	Detects open entries; assigns a pipeline_session_id; dispatches RTL then re-verification in sequence; enforces a configurable cap (default 3, via pipeline_config.max_cross_domain_iterations); scopes divergence checks to the current session; archives resolved entries on signoff; escalates via pending_approval if cap exceeded.

Domain Rules

fix_request Schema (authoritative)

All entries in design_state.fix_requests[] must conform to this schema. The machine-readable companion is docs/design_state.schema.json (JSON Schema Draft 2020-12), which CI validates fixtures against — it is the authoritative encoding of the enums, required fields, and the failure_class → retry_strategy map below.

{
  "id": "fr_<pipeline_session_id>_<YYYYMMDD>_<HHMMSS>_<seq>",
  "created_at": "<ISO-8601>",
  "updated_at": "<ISO-8601>",
  "created_by": "verification-orchestrator | formal-orchestrator",
  "failure_class": "functional | protocol | coverage_gap | formal_cex",
  "retry_strategy": "refine",
  "test_name": "<directed test or property name>",
  "property_or_assertion": "<assertion id or null>",
  "seed": 0,
  "waveform_path": "<path or null>",
  "log_path": "<path or null>",
  "suspected_rtl": {
    "module": "<module name>",
    "signal": "<signal or null>"
     
      
  
   
   
   
   
   
   
   

Reserved field — route_to (optional). The schema accepts an optional route_to string naming the servicer domain for a fix (default: rtl-design). It is forward-compatible scaffolding only: the pipeline-orchestrator currently always dispatches the RTL orchestrator (dispatch_to_producer), so producers need not set it. It exists so multi-servicer dispatch can be added later without a schema migration, mirroring the analog pipeline.

rtl_response (populated by rtl-design-orchestrator on close):

{
  "fixed_at": "<ISO-8601>",
  "diff_summary": "<one-paragraph description of changes>",
  "files_changed": ["rtl/path.sv"],
  "commit_ref": null
}

fix_request.history[] — one entry per state transition:

{
  "timestamp": "<ISO-8601>",
  "agent": "<agent name>",
  "from_status": "<previous status>",
  "to_status": "<new status>",
  "note": "<optional one-liner>"
}

Ownership rules

• rtl-design-orchestrator owns the open→claimed and claimed→fixed|abandoned transitions.
• Only the rtl-design-orchestrator sets status=claimed→fixed or claimed→abandoned.
• Only the pipeline-orchestrator sets cross_domain_iteration_count, pipeline_session_id, pipeline_config, and moves resolved entries to archive_fix_requests[].
• Domain orchestrators may set pending_approval exclusively with type: "checkpoint" at their own sign-off stage; type: "escalation" remains the sole responsibility of the pipeline-orchestrator.
• approved_checkpoints[] is written by the user (or by an orchestrator executing an explicit approval instruction) and read by all orchestrators.
• All agents may append to fix_request.history[] but must not overwrite each other's entries.

Iteration cap

cross_domain_iteration_count in design_state.json tracks the total number of verification↔RTL dispatch cycles for the current pipeline session. The cap is controlled by pipeline_config.max_cross_domain_iterations (default: 3 if absent). The orchestrator treats the cap as "reaches or exceeds" (use >= max_cross_domain_iterations), writing a pending_approval entry and exiting as soon as the iteration count is equal to or greater than the cap, preventing an off-by-one extra cycle before escalation:

{
  "pending_approval": {
    "type": "escalation",
    "stage": null,
    "agent": "pipeline-orchestrator",
    "reason": "fix_request loop exceeded 3 cross-domain iterations",
    "fix_request_id": "<id>",
    "last_summary": "<last rtl_response.diff_summary>",
    "requires_user": true
  }
}

The user must review the escalation, manually fix the RTL or adjust the testbench, then clear pending_approval (set to null) and reset cross_domain_iteration_count to 0 before invoking the pipeline-orchestrator again. Optionally increase pipeline_config.max_cross_domain_iterations if more iterations are warranted.

Pipeline session fields

Top-level fields in design_state.json managed by the pipeline-orchestrator and related infrastructure:

• pipeline_session_id ("ps_<YYYYMMDD>_<HHMMSS>" or null): identifies the active pipeline run. Set on entry, cleared (set to null) on successful signoff. Scopes divergence checks and archival to the current session — entries from prior sessions are ignored by the divergence check.
• pipeline_config (object): user-tunable pipeline settings. Written with defaults on first run; never overwritten if already present.
  • max_cross_domain_iterations (integer, default 3): the iteration cap for the feedback loop.
  • checkpoints (array of strings, default []): stage names that require human approval before the producing orchestrator may declare sign-off. Empty array ⇒ fully autonomous (preserves today's behavior). Example default: ["arch_signoff", "rtl_signoff", "signoff"]. Written by the user; domain orchestrators read but never overwrite this field.
• approved_checkpoints[]: list of stages the user has approved. Schema per entry: { "stage": "<stage name>", "approved_at": "<ISO-8601>", "approved_by": "user" }. Written by the user (or by an orchestrator acting on an explicit approval instruction); read by all orchestrators at their sign-off gate check.
• archive_fix_requests[]: resolved entries from completed pipeline sessions, moved here by the pipeline-orchestrator on successful signoff. Same schema as fix_requests[]. Never written by domain orchestrators.

Constraints Schema (authoritative)

design_state.constraints is the single source of truth for design-intent parameters across all domain orchestrators. Defined once here; domain SKILL.md files reference these keys and document their default fallbacks.

"constraints": {
  "clock": {
    "clk_mhz": null,
    "clk_uncertainty_ps": null
  },
  "pvt_corners": [
    { "name": "ss_setup", "process": "SS", "voltage_v": null, "temp_c": null, "checks": ["setup"] },
    { "name": "ff_hold",  "process": "FF", "voltage_v": null,     
  
   
     
     
     
     
     
     
  
   
     
     
     
  
   
     
     
     
     
          
  
   
     
     
     
     
     
     
     
  
   
     
     
     
     
     
     
  
   
     
     
     
  
   
     
     
     
     
     
     
  
   
     
     
     
  

Non-null values are documented defaults matching current hardcoded SKILL.md literals. null values must be supplied by the user for constraint-bearing domains.

Required vs. optional constraints

Constraint key	Required by (hard-fail if missing/null)
clock.clk_mhz	architecture, rtl-design, synthesis, sta, pd, soc, fpga
area.area_um2	architecture, synthesis, pd
power.power_mw	architecture, synthesis, pd
pvt_corners (≥1 entry with non-null voltage_v and temp_c)	sta, pd
hls.target_ii or hls.target_latency_cycles (at least one non-null)	hls

All other keys are optional — absent keys fall back to the schema default with a WARN.

Stage-entry constraint validation rule

Every constraint-bearing domain orchestrator applies this rule at the first stage that consumes design constraints. Skip entirely when invoked in fix-request-servicing mode (a fix_request.id was passed in the prompt):

1. Read design_state.constraints. Treat missing key as {}.
2. For each key in this domain's required set: if missing or null, perform atomic RMW — set pending_approval:

   {
     "type": "constraint_gap",
     "stage": "<entry stage name>",
     "agent": "<this-orchestrator>",
     "reason": "required constraint <key> missing from design_state.constraints",
     "fix_request_id": null,
     "last_summary": "<comma-separated list of missing keys>",
     "requires_user": true
   }
   

   Append a history[] entry: decision: "escalate", confidence: "high", failure_class: "spec_gap", suggested_next_step: "escalate", constraint_ref: "<missing key>". Print the gate message and halt.
3. For optional absent constraints: use the schema default, continue, and include a fallback note in the stage's history reason field.

Resume path: populate the missing constraint(s) in design_state.constraints, set pending_approval = null, and re-invoke the orchestrator.

Decision tagging via constraint_ref

In every history[] entry emitted after a stage that evaluates QoR against a constraint, set constraint_ref to the primary constraint key compared using dot-path notation (e.g. "timing.wns_ns_target", "clock.clk_mhz", "area.utilization_pct_max"). Comma-separate if a stage gates on multiple keys. All other history entries retain constraint_ref: null.

Decision tagging via retry_strategy

In every history[] entry, set retry_strategy to the value mapped from the entry's failure_class per the Failure Classification & Retry Strategy section. Entries with failure_class: "none" (PASS, await_approval) use retry_strategy: "none". This field is the strategy label read by the pipeline-orchestrator alongside confidence and suggested_next_step.

Failure Classification & Retry Strategy

Every failure is categorised so recovery is determined programmatically rather than by prose. Two fields work together on each history[] entry:

• failure_class — what went wrong (the existing 10-value enum, unchanged).
• retry_strategy — how to recover, derived deterministically from failure_class.

retry_strategy ∈ none | regenerate | refine | escalate:

• regenerate — discard the faulty artifact and re-run the generating stage from a clean slate using the error log as context (malformed/invalid output: tool crash, DRC/LVS, broken connectivity). Concrete action is typically retry_stage or loop_back_to:<generating stage>.
• refine — keep the artifact and re-run the stage targeting a specific identified defect with detailed feedback (failing test + waveform, timing path, coverage hole, violated interface). Iterative, not from scratch. Action is typically loop_back_to:<stage>, usually carrying a fix_request.
• escalate — halt and request human input; the result cannot be improved automatically (ambiguous spec) or a budget/cap was hit. Action is escalate / abandon.
• none — no failure (PASS or await_approval); pairs only with failure_class: "none".

retry_strategy is the strategy label; suggested_next_step remains the concrete action. They are complementary, not redundant.

Mapping (authoritative — failure_class → default retry_strategy)

failure_class	retry_strategy	rationale	legacy alias
none	none	no failure	—
functional	refine	re-run rtl_coding with failing test + waveform	verification_failure
timing	refine	re-run optimisation targeting failing paths	—
power_area	refine	re-run targeting the budget overage	—
coverage_gap	refine	add targeted stimulus to close holes	—
connectivity	refine	re-run targeting the violated interface/connection	interface_mismatch
drc_lvs	regenerate	re-run place/route from a clean state	—
tool_error	regenerate	re-run the same stage from scratch (≡ retry_stage)	invalid_rtl
spec_gap	escalate	ambiguous/missing spec — needs user clarification	incomplete_spec
resource_limit	escalate	iteration cap / memory exceeded — human decision	—

The "legacy alias" column reconciles the four classes proposed in an earlier draft (invalid_rtl | verification_failure | interface_mismatch | incomplete_spec) onto the live enum — no separate taxonomy is introduced. Producers that emit a fix_request (verification, formal) set its retry_strategy to refine (all their classes map to refine).

Actionable escalation guidance

Whenever retry_strategy resolves to escalate or a max-iteration cap is hit, the pending_approval.reason must state both the failure_class and a plain-language description of what the user must supply to unblock the flow, e.g.:

• spec_gap → "spec_gap: clarify — provide the intended <behaviour/value>."
• resource_limit → "resource_limit: loop cap (N) reached on — relax the constraint, raise the cap, or accept current QoR."

format_version

design_state.json version tiers (each tier is a superset of the previous):

• "1.1": fix_requests[] and cross_domain_iteration_count present.
• "1.2": history[] entries carry standardized confidence, failure_class, and suggested_next_step fields.
• "1.3": pipeline_config.checkpoints, approved_checkpoints[], pending_approval.type/stage/agent present; per-stage history[] entries (one entry per completed stage, not just one terminal entry per run).
• "1.4": constraints object present (authoritative nested schema defined in the Constraints Schema section); stage-entry constraint validation; pending_approval.type: "constraint_gap".
• "1.5": every history[] entry carries retry_strategy (none | regenerate | refine | escalate), derived from failure_class via the mapping in the Failure Classification & Retry Strategy section; escalations include failure_class + actionable guidance in pending_approval.reason.

All orchestrators must:

• Upgrade to "1.5" if absent or currently "1.0", "1.1", "1.2", "1.3", or "1.4"; never downgrade.
• All prior-version requirements are subsumed by "1.5" — no separate upgrades required.
• Treat missing fix_requests or cross_domain_iteration_count as [] / 0.
• Treat missing confidence, failure_class, or suggested_next_step in history entries as null for backward compatibility.
• Treat missing retry_strategy in history entries as derivable from failure_class via the mapping (none ⇒ none) for backward compatibility.
• Treat missing pipeline_config.checkpoints as [] (no checkpoints — fully autonomous).
• Treat missing approved_checkpoints as [].
• Treat missing pending_approval.type as "escalation" for backward compatibility.
• Treat missing constraints as {} — apply schema defaults for all optional keys; halt on first missing required key per the Constraints Schema section.

Approval Checkpoints

Proactive human-in-the-loop gates at configurable stage boundaries. Orthogonal to the failure-driven pending_approval (type escalation) set by the pipeline-orchestrator.

Checkpoint configuration

{
  "pipeline_config": {
    "checkpoints": ["arch_signoff", "rtl_signoff", "signoff"]
  },
  "approved_checkpoints": [
    { "stage": "arch_signoff", "approved_at": "<ISO-8601>", "approved_by": "user" }
  ]
}

Default: checkpoints: [] → no gates, fully autonomous (backward compatible).

Gate logic (applied by every domain orchestrator at its sign-off stage)

Before setting the domain's signoff=true and writing it to design_state.json:

1. Skip the gate in fix-request-servicing mode: if a fix_request.id was passed in the prompt (invoked by meta to repair a bug), skip the checkpoint check entirely. Checkpoints gate forward design progression, not the automated verify↔RTL repair loop.

2. Read pipeline_config.checkpoints from design_state.json. If the orchestrator's sign-off stage name appears in the list and does not appear in approved_checkpoints[].stage:

   • Atomic RMW: set pending_approval:

     {
       "type": "checkpoint",
       "stage": "<sign-off stage name>",
       "agent": "<this-orchestrator>",
       "reason": "checkpoint <stage> requires human approval before proceeding",
       "fix_request_id": null,
       "last_summary": "<QoR one-liner>",
       "requires_user": true
     }
     

   • Append a history[] entry with decision: "await_approval", confidence: "high", failure_class: "none", suggested_next_step: "escalate", reason: "checkpoint <stage> requires human approval".
   • Do not set signoff=true. Print the gate message to the user and halt.

Resume paths

The user may resume a gated orchestrator by either:

• Manual edit: append { "stage": "<stage>", "approved_at": "<ISO-8601>", "approved_by": "user" } to approved_checkpoints[] and set pending_approval=null in design_state.json, then re-invoke the orchestrator.
• Approval instruction: invoke the orchestrator with "approve checkpoint <stage>" in the prompt — the orchestrator performs the approved_checkpoints[] append and pending_approval clear (atomic RMW) itself, then continues.

pending_approval type-awareness (pipeline-orchestrator)

The pipeline-orchestrator's detect_open_fix_requests halts on any non-null pending_approval (conservative — a domain checkpoint also blocks meta dispatch). It prints a type-specific message:

• type: "checkpoint": "Checkpoint <stage> is awaiting human approval (set by <agent>). Approve or skip to continue."
• type: "escalation": (existing message) "Fix-request loop escalation — review required."
• type: "constraint_gap": "Stage <stage> is missing required constraint(s) (set by <agent>). Populate design_state.constraints and clear pending_approval to continue."

Per-stage history trace (format_version 1.3)

Every domain orchestrator appends one history[] entry after each internal stage completes (PASS, FAIL, WARN), not just at session end. The last entry written is the terminal entry read by the pipeline-orchestrator's decision table. At format_version 1.5 the entry carries a retry_strategy field derived from failure_class (10-field schema). This enables post-run audits without replaying the full conversation.

Programmatic branching on standardized history[] fields

After a domain orchestrator completes, the pipeline-orchestrator reads the terminal history[] entry to make retry/escalate decisions without string-parsing prose. Decision table (evaluated in order):

confidence	failure_class	retry_strategy	suggested_next_step	Pipeline action
any	resource_limit	escalate	any	Escalate via pending_approval — cap exceeded
low	any	any	escalate	Escalate — result unreliable, human review required
low	any	any	any (not escalate)	Escalate — low confidence overrides any retry intent
any	tool_error	regenerate	retry_stage	Re-dispatch the same orchestrator once; if still tool_error, escalate
any	drc_lvs | connectivity	refine	loop_back_to:<stage>	Re-dispatch the generating orchestrator from a clean slate with the error log
any	functional | coverage_gap	refine	escalate	Append new fix_request and loop back via RTL orchestrator
any	timing | power_area	refine	loop_back_to:<stage>	Re-dispatch targeting the violating path/budget with QoR feedback
any	spec_gap	escalate	escalate	Escalate — ambiguous spec; reason states what the user must clarify
high | medium	none	none	proceed	Advance to next stage / signoff

retry_strategy is the deterministic map of failure_class (see the Failure Classification & Retry Strategy section) and serves as a coarse pre-filter; confidence and suggested_next_step still refine the final action. Precedence is preserved: resource_limit and low confidence escalate regardless of the mapped strategy. For any combination not in the table, apply the most conservative matching rule (prefer escalate over retry). When the action is an escalation, pending_approval.reason must carry the failure_class and actionable guidance (see Actionable escalation guidance). Programmatic branches must read from the history entry's structured fields — do not re-derive intent from reason (free-text, for humans only).

Dispatch pattern (pipeline-orchestrator)

Sequential dispatch — never parallel:

1. RTL orchestrator (fix the bug) — block until complete.
2. Verification or formal orchestrator (validate the fix) — block until complete.

Spawn form for the Agent/Task tool:

• RTL: subagent_type: chip-design-rtl:rtl-design-orchestrator
• Verification: subagent_type: chip-design-verification:verification-orchestrator
• Formal: subagent_type: chip-design-formal:formal-orchestrator

Always pass the fix_request.id in the subagent prompt so the child can locate its work item without scanning the whole array.

V2 extension points (not wired in V1)

• Architecture↔RTL refinement loop: architecture.refinement_needed=true could trigger an arch re-run. The fix_request schema is intentionally producer-agnostic; only created_by would need a new value (architecture-orchestrator).
• Formal property-bug routing: failure_class=formal_cex with suspected_owner=formal would route to the formal orchestrator instead of RTL. Not implemented in V1.
• LEC unmatched-points loop: lec_run: unmatched points in formal-orchestrator.md is intentionally not connected to the fix_request protocol in V1. LEC failures are netlist↔RTL mismatches introduced at synthesis — the correct consumer is synthesis-orchestrator, not rtl-design-orchestrator. Deferred to V2.

QoR Metrics

• cross_domain_iteration_count — number of RTL↔verify dispatch cycles (target: ≤ 2 for clean designs)
• time_to_signoff — wall-clock time from first open fix_request to verification_status.signoff=true
• escalation_rate — fraction of design sessions that hit the 3-iteration cap (target: < 10%)
• fix_request_abandonment_rate — fraction of fix_requests that reach status=abandoned (target: 0%)

Output Required

• design_state.json with all fix_requests[] entries at terminal status (fixed or abandoned)
• design_state.json with cross_domain_iteration_count updated
• memory/meta/experiences.jsonl entry for this run
• A console summary to the user: which fix_requests were processed, how many iterations, and the outcome (converged / escalated / no open requests)