forge-framing

Framing Methodology

Loaded during the Frame phase. Governs how the Reasoner decomposes problems into tractable decision structures.

When to use

Load this skill at the start of Phase F (Frame). Use it to extract the core decision question, classify the decision type (TRADE-OFF, FEASIBILITY, ROOT-CAUSE, CONFLICT-RESOLUTION, CONSTRAINT-SATISFACTION, RISK-ASSESSMENT), map all hard and soft constraints, set the deliberation depth score (1–9), and write the success criteria checklist. Do not proceed to Phase O (Observe) until all five steps are complete. Do not use for the Reason, Gate, or Emit phases.

---

Memory pre-flight (Frame — mission intake)

Before any decomposition work begins, call CRYSTALIUM recall to surface prior verdicts, known constraints, and recurring deliberation patterns for this project:

mcp__crystalium__recall(
  scope    = { project: <cwd-project>, agent_class_visibility: "forge" },
  query    = <the decision question being framed>,
  k        = 5,
  layers   = ["semantic", "episodic", "procedural"]
)

Fold relevant hits (prior verdicts on related decisions, reversal conditions that have already fired, recurring constraint patterns) into the evidence inventory before Step 1 below. Prior deliberation outputs surface at T1 episodic, making prior FORGE verdicts directly reusable.

Graceful skip: if mcp__crystalium__* tools are unavailable (CRYSTALIUM not installed), proceed without memory — never hard-fail. FORGE is EIIS-standalone- conformant and works without CRYSTALIUM.

See PERSONA.md §"Memory pre-flight" for the always-loaded note. See SPEC.md §9 for the full memory protocol summary.

---

Problem Decomposition Protocol

Step 1: Extract the Core Decision

Strip away narrative to find the atomic question. Apply the Specificity Test:

If two competent engineers could interpret this question differently and produce contradictory verdicts, the framing is insufficiently specific.

Transform vague questions into precise ones:

Vague	Specific
"Should we use microservices?"	"Given our team of 4, a monolith at 50K LOC, and 3 independent deployment cadences, does decomposing services X, Y, Z reduce deployment risk enough to justify the operational overhead?"
"Is this architecture scalable?"	"Can this architecture handle 10x current load (defined as N req/s) within the existing infrastructure budget of $M/month, without exceeding p99 latency of Xms?"
"What caused this outage?"	"Which component in the request path between ingress and database was the first to degrade, and what trigger pushed it past its failure threshold?"

Step 2: Classify the Decision Type

Each type activates different reasoning patterns in the Deliberation phase:

Type	Core Question Shape	Reasoning Pattern
TRADE-OFF	"X vs Y given constraints C"	Multi-criteria scoring with explicit weights
FEASIBILITY	"Can X be achieved under constraints C?"	Constraint enumeration → bottleneck identification → existence proof or counterexample
ROOT-CAUSE	"Why did X happen?"	Causal chain reconstruction → elimination of confounders
CONFLICT-RESOLUTION	"Agents/stakeholders disagree on X"	Position extraction → shared-ground identification → arbitration
CONSTRAINT-SATISFACTION	"Find solution satisfying C1, C2, ... Cn"	Constraint graph → feasibility region → optimization within region
RISK-ASSESSMENT	"What can go wrong with X?"	Failure mode enumeration → likelihood × impact scoring → mitigation mapping

Step 3: Map Constraints

Build the constraint table. Every constraint has a source and a hardness level:

| ID | Constraint | Hard/Soft | Source | Risk if Violated |
|----|-----------|-----------|--------|-----------------|
| C1 | p99 latency < 200ms | Hard | SLA contract | Service credit penalties |
| C2 | No new infrastructure spend | Soft | Q3 budget guidance | Deferral to Q4 |
| C3 | Must work with PostgreSQL 14 | Hard | Ops team mandate | Migration blocked |

Hard constraints eliminate options. Soft constraints influence scoring but don't eliminate.

Step 4: Set Deliberation Depth

Score the problem on three dimensions (1–3 each):

Dimension	1 (Simple)	2 (Standard)	3 (Deep)
Ambiguity	Clear right answer exists	2–3 viable paths	Many viable paths, unclear winner
Reversibility	Easy to undo	Moderate cost to reverse	Irreversible or very costly
Blast radius	Single component	Multiple components	System-wide or organizational

Total 3–4 → Simple (1 reasoning pass). Total 5–7 → Standard (2 passes). Total 8–9 → Deep (3 passes, extended token budget).

Step 5: Declare Success Criteria

Before reasoning begins, write down what "done" looks like:

## Success Criteria
- [ ] The verdict answers: [restate the specific question]
- [ ] All hard constraints addressed with pass/fail
- [ ] At least 3 alternatives evaluated with rejection reasons
- [ ] Reversal conditions stated
- [ ] Confidence score defensible
- [ ] Requester can act without further clarification

---

Evidence Inventory Protocol

Catalog every piece of provided context before reasoning:

## Evidence Inventory

| ID | Source | Type | Relevance | Reliability |
|----|--------|------|-----------|-------------|
| E1 | ATLAS scout-report.md | Scout report | Direct | H — evidence-anchored |
| E2 | User-provided benchmark data | Performance data | Direct | M — methodology unstated |
| E3 | Team Slack discussion | Stakeholder position | Contextual | L — informal, may be outdated |
| E4 | SPECTRA spec v2 | Planning artifact | Direct | H — verified through SPECTRA gates |

Reliability tiers:

• H (High) — Evidence-anchored artifacts, verified test results, official documentation
• M (Medium) — Reasonable but unverified: benchmarks without methodology, expert opinions, dated documentation
• L (Low) — Informal, secondhand, or potentially outdated sources

Claims built on L-reliability evidence must carry [ASSUMPTION] markers.

---

Anti-Patterns in Framing

Anti-Pattern	Signal	Remedy
Binary framing	"Should we do X or not?"	Expand to ≥3 options: do X, do Y, do nothing, do X-lite
Solution-first framing	Question presupposes the answer	Re-frame around the problem, not the solution
Scope creep during framing	Sub-questions multiply beyond the original ask	Constrain to the declared question; log expansions as follow-up verdicts
Missing stakeholders	Constraints exist that nobody stated	Ask: "Who else is affected? What would they say?"
Phantom constraints	Assumed constraints that nobody actually mandated	Verify every constraint has a named source

---

Reasoner — Framing Skill