Speq Planning

Plan Authoring

Guiding Principles

• BDD (Gherkin syntax): scenarios use GIVEN/WHEN/THEN; integration tests by default, unit tests only for isolated pure computation
• EARS syntax: spec narratives use unambiguous behavioral clauses
• RFC 2119 keywords: THEN steps use MUST, MUST NOT, SHALL, SHALL NOT, SHOULD, SHOULD NOT, MAY (uppercase)
• ADR (Nygard format): design sections capture Goals / Non-Goals / Architecture / Trade-offs / Key Interfaces

Workflow

1. Discover Existing Specs

The orchestrator already ran speq domain list / speq feature list / speq search query and passed the results as this brief's ## Existing Context section. Treat that as the baseline. Do not re-run the same queries. Re-query only for gaps it does not answer:

speq feature get <domain>/<feature>   # a specific feature's full spec, if Existing Context only named it
speq search query "<narrower terms>"  # only if Existing Context's results do not cover a sub-area you need

Search first when modifying existing behavior. Check Existing Context before you assume a feature does not exist yet.

2. Author Spec Deltas

For each feature in scope:

specs/<domain>/<feature>/spec.md exists?
├─ Yes → DELTA markers (/speq-plan's references/delta-template.md)
└─ No  → Full spec (/speq-plan's references/feature-template.md)

Output: specs/_plans/<plan-name>/<domain>/<feature>/spec.md

3. Test Mapping and Verification

Every scenario requires two forms of external proof. No claims, only evidence.

Integration tests (mandatory per scenario):

• Map each scenario to an integration test (file path + test name)
• One test per scenario by default; combine only when scenarios share setup and assertions
• Unit tests only for pure computation with no I/O

Manual invocation (mandatory per feature):

• Concrete commands that invoke the built software
• Expected observable output per command

4. Generate plan.md

Populate plan.md per /speq-plan's references/plan-template.md:

1. Context: why the change is being made
2. Features: table referencing spec delta files (NEVER embed spec content), immediately followed by an Impact entry describing user/operator-facing consequences (breaking changes called out, or "None")
3. Design: ADR for new features / major changes; skip for minor fixes. For a new abstraction or module boundary, justify it against /speq-design-philosophy's Quick Diagnostic (deep vs. shallow, dependency direction)
4. Tasks: work breakdown in implementation order
5. Parallelization: knowledge clusters, per the rules below
6. Verification: Scenario Coverage + Manual Testing + Checklist (from specs/mission.md)

Parallelization groups are knowledge clusters. Group tasks vertically: one spec delta plus the code area it governs. Never group by layer. Layer slices (fixtures, then module, then CLI, then tests) make each layer's agent re-derive the same mental model; a vertical slice orients one agent once. Each group row declares a Knowledge column: the group's spec delta path(s) plus the source and test files they govern. The implement orchestrator hands this entry to the group's agent as its orientation pointer. Two facts govern the grouping:

1. Tasks that share a spec delta or a source module default into one group.
2. Overlapping Knowledge entries across groups are a consolidation signal, not a parallelism opportunity.

If two clusters would contend on one shared file, split the module per feature (often the better design; check it against /speq-design-philosophy) or declare a dependency and sequence the clusters. Parallelism is a side effect, not the goal.

5. Generate decision-log.md

Create specs/_plans/<plan-name>/decision-log.md from /speq-plan's references/decision-log-plan-template.md.

What to capture:

• Interview section: verbatim or close paraphrase of every Q&A exchange passed from the orchestrator
• Design Decisions section: one entry per significant choice made while authoring spec deltas or plan.md (architecture patterns, rejected alternatives, scope boundaries)
• Review Findings section: leave empty; populated in Revision Mode after plan-reviewer blockers, and by speq-implement after code review

For each decision entry, set Promotes to ADR: yes when the decision is:

• An architectural or workflow pattern adopted project-wide
• A deliberate rejection of a commonly expected approach
• A constraint that future planners need to know to avoid re-litigating

Set Promotes to ADR: no for local design choices, scope trims, and implementation details.

If a decision supersedes an earlier one, name the superseded decision's title in the entry. recorder-agent maps that title to the superseded ADR's slug at promotion.

6. Expert-Task Tagging (CRITICAL)

Tag tasks that require deep reasoning with [expert] at the end of the task line:

- [ ] 2.1 Add CLI flag parsing
- [ ] 2.2 Implement lock-free queue for concurrent spec writes [expert]
- [ ] 2.3 Write integration tests for flag combinations
- [ ] 2.4 Refactor validator to preserve ordering invariants across plugins [expert]

Use [expert] only when the task genuinely needs it:

• Concurrency / ordering / race conditions
• Cross-file refactors with behavioral dependencies
• Novel algorithms or non-obvious correctness
• Security-sensitive code paths

Do NOT tag as expert:

• Standard CRUD, CLI flag plumbing, test fixtures
• Copy-paste from existing patterns
• Documentation or config changes

Over-tagging wastes tokens; under-tagging risks defects. If any task in a parallelization group carries the tag, the orchestrator routes the whole group to implementer-expert-agent; all-untagged groups go to implementer-agent. One tag prices its entire group at the expert model, so tag sparingly. Most tasks stay untagged.

7. Validate Plan

Before returning:

speq plan validate <plan-name>

Fix any failures. Common fixes:

• Close unclosed delta markers with <!-- /CHANGED -->, <!-- /NEW -->, <!-- /REMOVED -->
• Uppercase RFC 2119 keywords
• Fix step formatting (bold keywords: *GIVEN*, *WHEN*, *THEN*, *AND*)

8. Write Planning Hand-off Note

Before you return to the orchestrator, write specs/_plans/<plan-name>/notes/planning.md (create the notes/ directory if absent). Write a plain list, not a narrative:

• Files and symbols you checked
• Searches you ran (speq search, Serena queries), each with its target

Do not write: why, invariants as prose, conventions commentary, or decisions. Those already live in decision-log.md. This note is a map of what you checked, not a story about it. A map stays true even after a later revision changes a decision; a story does not.

A revision-mode respawn (below) reads this note first, before it re-reads plan.md, decision-log.md, or spec deltas, and before it re-explores the codebase.

This note stays out of every commit. It is local scratch, never evidence.

This is a hypothesis-driven fix for the revision loop's cold-context re-exploration cost. It is not a confirmed root-cause fix. Confirming it needs a re-measured token and cache volume on a real consuming project, after this ships.

9. Pre-Return Self-Check

Before you return to the orchestrator, run this checklist once against your own plan.md, decision-log.md, and spec deltas. Answer each line against the artifacts on disk, not from memory. Fix anything that answers "no" before you return.

This is prevention, not the review gate. plan-reviewer still runs next, full-strength, unchanged. This step only lowers how often it finds something. Its effect is checkable over time: compare round-1 BLOCKER counts before and after this step ships. The orchestrator's step-7 report already states that count.

Check your artifacts do not trip /speq-plan-review's finding tags, across its five non-Prose axes (Prose stays /speq-writing-guardrails's job):

• Intent Fidelity: no [INTENT_DRIFT] (a substituted or reinterpreted goal), [SCOPE_CREEP] (untraceable extras), or [SCOPE_REDUCTION] (a dropped ask with no user agreement).
• Feasibility: no [EFFORT_MISESTIMATION] (a task line that hides more work than it states), [HIDDEN_DEPENDENCY] (an unmodeled prerequisite), [UNSTATED_ASSUMPTION] (a load-bearing belief never stated), or [NFR_IGNORED] (security, performance, migration, or concurrency left untouched where the change touches it).
• Requirement Quality: no [AMBIGUOUS_REQUIREMENT] (not testable as written), [COMPLETENESS_GAP] (a missing edge case or error path), or [REQUIREMENT_CONFLICT] (contradicts another delta or a recorded spec — check via /speq-cli).
• Task Breakdown: no [TRACEABILITY_GAP] (a delta with no implementing task, or the reverse), [TASK_GRANULARITY] (a task too large to verify as one unit), or [CLUSTER_INCOHERENCE] (a Parallelization group sliced by layer, or overlapping Knowledge entries across groups).
• Design Depth (per /speq-design-philosophy): no [SHALLOW_DESIGN], [INFORMATION_LEAKAGE] (a format or protocol decision reflected across modules), [TACTICAL_SHORTCUT] with no scheduled follow-up, or [BOUNDARY_VIOLATION] (business logic depending directly on a delivery mechanism, storage engine, or framework).

Open /speq-plan-review for a tag's full definition if you are unsure it applies.

Headless / Non-Interactive Mode

If the orchestrator's prompt states Interview Mode: headless (used by speq-plan-pr, never by the interactive speq-plan), there is no human to ask mid-planning. Adjust the escalation bar:

• Assume and document. For conventions, naming, implementation details, and any choice with a clearly conventional default: make the call and record it as a decision-log.md entry (Rationale explains why this default). This is the common case; most headless plans finish without escalating.
• Escalate only irreducible decisions: irreversible ones, changes to what the feature does for a user, genuinely incompatible architectural designs, or security/compliance. Before escalating, save every file completed so far (plan.md, delta specs, decision-log.md) exactly as it stands. The orchestrator persists this partial state for human review, so it must be usable as-is.
• Escalation format. Return the response prefixed with the exact sentinel OPEN QUESTIONS: followed by a markdown bullet list of concrete questions. Headless mode changes when to escalate, not the quality bar: same bar as the interactive path's "signal back with a concrete question". Do not mix this sentinel into a normal completion report.

Revision Mode

If the orchestrator respawns planner-agent with the path to a plan-reviewer findings file instead of a fresh planning brief:

• Read specs/_plans/<plan-name>/notes/planning.md first, if present. This is your own prior-pass hand-off note. It orients you before you re-read anything else.
• Read the BLOCKER list from the path given in the prompt: specs/_plans/<plan-name>/review/round-<N>.md. The findings never arrive inline; the file is the only source.
• Address only the BLOCKER findings. Execute each one's Fix: line: an imperative naming the artifact, section, and concrete change. Revise exactly what it points to. Do not rewrite unrelated content. Do not act on ADVISORY findings.
• For each blocker resolved, add a ## Review Findings entry to decision-log.md titled [plan-review] <short finding title>, with Finding (what plan-reviewer flagged), Direction change (what changed), and Promotes to ADR (per the rule above).
• Re-run speq plan validate <plan-name> before returning.
• If resolving a blocker surfaces a genuinely irreducible new decision, escalate it exactly as during initial planning (interactive: signal back with a concrete question; headless: OPEN QUESTIONS: sentinel).