Epic execution profile
The Epic execution profile is an opt-in overlay on core Forge for teams that delegate L3 use-case slices to agents under OpenSpec acceptance. Humans Charge and specify Epics (intent + observable acceptance); agents…
Guide · Updated · Source
Introduction
The Epic execution profile is an opt-in overlay on core Forge for teams that delegate L3 use-case slices to agents under OpenSpec acceptance. Humans Charge and specify Epics (intent + observable acceptance); agents execute L1–L2 work inside a Charged Epic as runs, not as Charge-visible WBS Task rows.
Dual profile. Core Forge keeps Ore → Ingot → Forge Spark → Charge (Sparks). Under the Epic execution profile, Charge lists Epics instead of Sparks. Spark is not removed from Forge — teams choose one delivery profile (or run both on different repos). Every retargeted doc says under the Epic execution profile when describing Epic-specific behavior.
When to adopt: A team is ready when most committed delivery is agent-executed L3 slices with reviewable diffs, one primary repo per Epic, and OpenSpec acceptance criteria the Assay gate can test.
Size gate (ready Epic)
An Epic is ready to Charge only when it passes the size gate:
| Gate | Requirement |
|---|---|
| One OpenSpec change | Exactly one OpenSpec change folder per Epic — one intent, one reviewable scope |
| Testable SHALLs | Capability spec → at least one observable SHALL / MUST → ≥1 Given / When / Then scenario — not implementation SHALLs (no file or class names in acceptance) |
| One primary repo | Execution and review stay in one primary repository; cross-repo work triggers L4.2 stop-and-ask (below) |
| Reviewable diff | Expected outcome fits a single PR or merge request a human can review in one sitting |
| Dual wiki (local) | dual wiki on the go: declared ## Dual wiki surfaces on the OpenSpec proposal have fresh local handbook HTML — overlay wiki_stale blocks Specify→Ready until loopback refresh; not Firebase publish (Spec Flow board) |
Epics that fail the gate stay in refinement (OpenSpec draft, scope split, or human L4 planning) — do not Charge until the gate passes.
Grain under the profile
| Grain | Role under the Epic execution profile |
|---|---|
Product Spark (M1) |
Release / Assay horizon — unchanged |
Epic (M1E3) |
Smallest committed unit = autonomy L3 (OpenSpec AC in; review/merge out) |
| Story / Task / Forge Spark | Not Charge work items; L1–L2 agent scratch only inside the Epic |
| Charge | Epics in play — a view, not a Kanban system of record |
| OpenSpec change | 1:1 with a ready Epic; Lite SHALLs + scenarios; tasks.md is non-binding |
| L4+ | Extra human gates (ADR, cross-repo L4.2); agents stop and ask |
| Assay | Still Product Spark scope; archive ≠ Released |
WBS Epic IDs (e.g. M1E3) align with committed Epics under this profile. Do not mint M1E3S2T4 rows on Charge for Charged Epics.
L1–L3 inside the Epic
Under the Epic execution profile, autonomy levels L1–L3 execute inside the Charged Epic boundary:
| Level | Inside Epic | Human gate |
|---|---|---|
| L1 | Single function / contract-bound patch | Approve branch / merge |
| L2 | Multi-file change-set without rearchitecture | Accept acceptance + merge |
| L3 | End-to-end use-case slice in one existing app (default Epic target) | OpenSpec acceptance in; review out |
The committed Epic maps to L3 use-case slice by default — see Autonomy levels § Epic execution profile. Agents may run L1 or L2 sub-work inside the Epic without new Charge rows.
Decompose runs, not backlogs: Agents break work into execution runs (classify → plan → apply → verify → repair → escalate). They do not maintain parallel WBS Story/Task backlogs or mint Charge rows for each run.
OpenSpec Lite and Full
| Profile | When | Acceptance bar |
|---|---|---|
| OpenSpec Lite (default) | Typical feature, UI, or internal behavior in one repo | One capability → observable SHALL → ≥1 scenario; tasks.md hints only |
| OpenSpec Full | API contracts, security-sensitive work, or cross-cutting platform change | Stricter change folder layout, explicit deltas, additional review gates per team OpenSpec pack |
Rules (Lite default):
- One change = one intent.
- Not implementation SHALLs — acceptance describes observable behavior, not source files.
- Humans approve OpenSpec acceptance before Charge; agents apply the change and prove scenarios in verify.
Full profile selection is a human decision at Epic specification time.
Charge of Epics
Under the Epic execution profile:
- Humans select ready Epics (size gate green, OpenSpec acceptance approved) into today's Charge.
- Charge remains a view on committed work — not a second board (Forge — major processes & flow maps). The Spec Flow board is a derived lens on WBS + OpenSpec +
forge/charge.md; drag writes Charge and change phase — it is not a parallel Charge SoT. - One Charged Epic may host many agent runs at L1–L3; runs are not Charge items.
- Daily sync confirms Epic progress, blockers, and escalation — not a Task-level standup on WBS rows.
Active Epics Charge statuses map to board columns: planned → Charged, in progress → Apply, done → Verify — see Spec Flow board § Charge status map.
Core Forge teams that Charge Sparks unchanged — see Daily operations.
L4.2 stop-and-ask
When work crosses L4 boundaries, agents stop and ask a human before proceeding:
| Trigger | Agent behavior |
|---|---|
| Cross-repo change (e.g. submodule bump, L4.2) | Stop; request human plan, ADR draft, or explicit approval |
| New component or platform pattern (L4.1+) | Stop; escalate to architecture gate |
| Security / compliance ambiguity | Stop; invoke Security Versona or human owner |
| OpenSpec Full required but only Lite present | Stop; request profile upgrade or scope reduction |
Under the Epic execution profile, L4+ is out of autonomous scope unless a human explicitly expands the Epic with new gates and acceptance.
Trivial L1 skip
Trivial one-liners (typo, comment, config knob with obvious test) may skip Epic ceremony when a classifier (human or agent policy) marks them L1 trivial:
- No new OpenSpec change folder required.
- No Epic Charge row required.
- Still subject to normal review / merge gates.
When in doubt, use an Epic — the skip is for noise reduction, not scope hiding.
Ember Log and decisions
Log Ember Log entries for Epic-scope or cross-cutting decisions only — not every L1 run inside an Epic. Operational run trace lives in machine records (Assay evidence, PR, OpenSpec verify output).
Related
| Topic | File |
|---|---|
| Spec Flow board (derived Kanban lens) | Spec Flow board |
| Autonomy L3 default for committed Epic | Autonomy levels |
| Naming: Epic, Charge dual-profile | Forge & planning — naming reference |
| Concept matrix + collisions | ForgeSDLC — concept map and term-collision register |
| Artifacts under profile | Forge — artifact and decision model |
| Core Spark delivery (default Forge) | Forge — major processes & flow maps · Forge SDLC |
| Bounded execution examples | Bounded execution examples |
Executive capsule
The Epic execution profile is an opt-in overlay: Charge lists Epics, OpenSpec acceptance defines L3 scope, and agents decompose runs inside the Epic — not WBS Task Charge rows. Core Forge keeps Spark → Charge. Maturity: defined.
Who this is for
Engineering leaders adopting agent-executed L3 delivery. Read the dual-profile note first; default Forge teams can ignore this page.
Trust boundary
Humans own Charge selection, OpenSpec acceptance, merge, and Assay. Agents execute inside declared Epic boundaries and stop and ask at L4.2+.
How to use this page
Confirm the size gate and OpenSpec bar before Charging an Epic; link execution policy to Autonomy levels for ladder enforcement.