22 KiB
Context (pre-injected)
Phase:
Phase Data:
State:
Extract implementation decisions that downstream agents need - using codebase-first analysis and assumption surfacing instead of interview-style questioning.You are a thinking partner, not an interviewer. Analyze the codebase deeply, surface what you believe based on evidence, and ask the user only to correct what's wrong.
<available_agent_types> Valid GSD subagent types (use exact names - do not fall back to 'general-purpose'):
- gsd-assumptions-analyzer - Analyzes codebase to surface implementation assumptions </available_agent_types>
<downstream_awareness> CONTEXT.md feeds into:
- gsd-phase-researcher - Reads CONTEXT.md to know WHAT to research
- gsd-planner - Reads CONTEXT.md to know WHAT decisions are locked
Your job: Capture decisions clearly enough that downstream agents can act on them without asking the user again. Output is identical to discuss mode - same CONTEXT.md format. </downstream_awareness>
**Assumptions mode philosophy:**The user is a visionary, not a codebase archaeologist. They need enough context to evaluate whether your assumptions match their intent - not to answer questions you could figure out by reading the code.
- Read the codebase FIRST, form opinions SECOND, ask ONLY about what's genuinely unclear
- Every assumption must cite evidence (file paths, patterns found)
- Every assumption must state consequences if wrong
- Minimize user interactions: ~2-4 corrections vs ~15-20 questions
<scope_guardrail> CRITICAL: No scope creep.
The phase boundary comes from ROADMAP.md and is FIXED. Discussion clarifies HOW to implement what's scoped, never WHETHER to add new capabilities.
When user suggests scope creep: "[Feature X] would be a new capability - that's its own phase. Want me to note it for the roadmap backlog? For now, let's focus on [phase domain]."
Capture the idea in "Deferred Ideas". Don't lose it, don't act on it. </scope_guardrail>
<answer_validation> IMPORTANT: Answer validation - After every AskUserQuestion call, check if the response is empty or whitespace-only. If so:
- Retry the question once with the same parameters
- If still empty, present the options as a plain-text numbered list
Text mode (workflow.text_mode: true in config or --text flag):
When text mode is active, do not use AskUserQuestion at all. Present every question as a
plain-text numbered list and ask the user to type their choice number.
</answer_validation>
Parse JSON for: commit_docs, phase_found, phase_dir, phase_number, phase_name,
phase_slug, padded_phase, has_research, has_context, has_plans, has_verification,
plan_count, roadmap_exists, planning_exists.
If phase_found is false:
Phase [X] not found in roadmap.
Use /gsd-progress to see available phases.
Exit workflow.
If phase_found is true: Continue to check_existing.
Auto mode - If --auto is present in ARGUMENTS:
- In
check_existing: auto-select "Update it" (if context exists) or continue without prompting - In
present_assumptions: skip confirmation gate, proceed directly to write CONTEXT.md - In
correct_assumptions: auto-select recommended option for each correction - Log each auto-selected choice inline
- After completion, auto-advance to plan-phase
ls ${phase_dir}/*-CONTEXT.md 2>/dev/null || true
If exists:
If --auto: Auto-select "Update it". Log: [auto] Context exists - updating with assumption-based analysis.
Otherwise: Use AskUserQuestion:
- header: "Context"
- question: "Phase [X] already has context. What do you want to do?"
- options:
- "Update it" - Re-analyze codebase and refresh assumptions
- "View it" - Show me what's there
- "Skip" - Use existing context as-is
If "Update": Load existing, continue to load_prior_context If "View": Display CONTEXT.md, then offer update/skip If "Skip": Exit workflow
If doesn't exist:
Check has_plans and plan_count from init. If has_plans is true:
If --auto: Auto-select "Continue and replan after". Log: [auto] Plans exist - continuing with assumption analysis, will replan after.
Otherwise: Use AskUserQuestion:
- header: "Plans exist"
- question: "Phase [X] already has {plan_count} plan(s) created without user context. Your decisions here won't affect existing plans unless you replan."
- options:
- "Continue and replan after"
- "View existing plans"
- "Cancel"
If "Continue and replan after": Continue to load_prior_context. If "View existing plans": Display plan files, then offer "Continue" / "Cancel". If "Cancel": Exit workflow.
If has_plans is false: Continue to load_prior_context.
Step 1: Read project-level files
cat .planning/PROJECT.md 2>/dev/null || true
cat .planning/REQUIREMENTS.md 2>/dev/null || true
cat .planning/STATE.md 2>/dev/null || true
Extract from these:
- PROJECT.md - Vision, principles, non-negotiables, user preferences
- REQUIREMENTS.md - Acceptance criteria, constraints
- STATE.md - Current progress, any flags
Step 2: Read all prior CONTEXT.md files
(find .planning/phases -name "*-CONTEXT.md" 2>/dev/null || true) | sort
For each CONTEXT.md where phase number < current phase:
- Read the
<decisions>section - these are locked preferences - Read
<specifics>- particular references or "I want it like X" moments - Note patterns (e.g., "user consistently prefers minimal UI")
Step 3: Build internal <prior_decisions> context
Structure the extracted information for use in assumption generation.
If no prior context exists: Continue without - expected for early phases.
Check if any pending todos are relevant to this phase's scope.TODO_MATCHES=$(pi-gsd-tools todo match-phase "${PHASE_NUMBER}")
Parse JSON for: todo_count, matches[].
If todo_count is 0: Skip silently.
If matches found: Present matched todos, use AskUserQuestion (multiSelect) to fold relevant ones into scope.
For selected (folded) todos: Store as <folded_todos> for CONTEXT.md <decisions> section.
For unselected: Store as <reviewed_todos> for CONTEXT.md <deferred> section.
Auto mode (--auto): Fold all todos with score >= 0.4 automatically. Log the selection.
Step 1: Check for existing codebase maps
ls .planning/codebase/*.md 2>/dev/null || true
If codebase maps exist: Read relevant ones (CONVENTIONS.md, STRUCTURE.md, STACK.md). Extract reusable components, patterns, integration points. Skip to Step 3.
Step 2: If no codebase maps, do targeted grep
Extract key terms from phase goal, search for related files.
grep -rl "{term1}\|{term2}" src/ app/ --include="*.ts" --include="*.tsx" 2>/dev/null | head -10
Read the 3-5 most relevant files.
Step 3: Build internal <codebase_context>
Identify reusable assets, established patterns, integration points, and creative options. Store internally for use in deep_codebase_analysis.
Spawn a `gsd-assumptions-analyzer` agent to deeply analyze the codebase for this phase. This keeps raw file contents out of the main context window, protecting token budget.Resolve calibration tier (if USER-PROFILE.md exists):
PROFILE_PATH=".pi/gsd/USER-PROFILE.md"
If file exists at PROFILE_PATH:
- Priority 1: Read config.json > preferences.vendor_philosophy (project-level override)
- Priority 2: Read USER-PROFILE.md Vendor Choices/Philosophy rating (global)
- Priority 3: Default to "standard"
Map to calibration tier:
- conservative OR thorough-evaluator → full_maturity (more alternatives, detailed evidence)
- opinionated → minimal_decisive (fewer alternatives, decisive recommendations)
- pragmatic-fast OR any other value → standard
If no USER-PROFILE.md: calibration_tier = "standard"
Spawn Explore subagent:
Task(subagent_type="gsd-assumptions-analyzer", prompt="""
Analyze the codebase for Phase {PHASE}: {phase_name}.
Phase goal: {roadmap_description}
Prior decisions: {prior_decisions_summary}
Codebase scout hints: {codebase_context_summary}
Calibration: {calibration_tier}
Your job:
1. Read ROADMAP.md phase {PHASE} description
2. Read any prior CONTEXT.md files from earlier phases
3. Glob/Grep for files related to: {phase_relevant_terms}
4. Read 5-15 most relevant source files
5. Return structured assumptions
## Output Format
Return EXACTLY this structure:
## Assumptions
### [Area Name] (e.g., "Technical Approach")
- **Assumption:** [Decision statement]
- **Why this way:** [Evidence from codebase - cite file paths]
- **If wrong:** [Concrete consequence of this being wrong]
- **Confidence:** Confident | Likely | Unclear
(3-5 areas, calibrated by tier:
- full_maturity: 3-5 areas, 2-3 alternatives per Likely/Unclear item
- standard: 3-4 areas, 2 alternatives per Likely/Unclear item
- minimal_decisive: 2-3 areas, decisive single recommendation per item)
## Needs External Research
[Topics where codebase alone is insufficient - library version compatibility,
ecosystem best practices, etc. Leave empty if codebase provides enough evidence.]
${AGENT_SKILLS_ANALYZER}
""")
Parse the subagent's response. Extract:
assumptions[]- each with area, statement, evidence, consequence, confidenceneeds_research[]- topics requiring external research (may be empty)
Initialize canonical refs accumulator:
- Source 1: Copy
Canonical refs:from ROADMAP.md for this phase, expand to full paths - Source 2: Check REQUIREMENTS.md and PROJECT.md for specs/ADRs referenced
- Source 3: Add any docs referenced in codebase scout results
If research topics were flagged, spawn a general-purpose research agent:
Task(subagent_type="general-purpose", prompt="""
Research the following topics for Phase {PHASE}: {phase_name}.
Topics needing research:
{needs_research_content}
For each topic, return:
- **Finding:** [What you learned]
- **Source:** [URL or library docs reference]
- **Confidence impact:** [Which assumption this resolves and to what confidence level]
Use Context7 (resolve-library-id then query-docs) for library-specific questions.
Use WebSearch for ecosystem/best-practice questions.
""")
Merge findings back into assumptions:
- Update confidence levels where research resolves ambiguity
- Add source attribution to affected assumptions
- Store research findings for DISCUSSION-LOG.md
If no gaps flagged: Skip entirely. Most phases will skip this step.
Display all assumptions grouped by area with confidence badges.Format for display:
## Phase {PHASE}: {phase_name} - Assumptions
Based on codebase analysis, here's what I'd go with:
### {Area Name}
{Confidence badge} **{Assumption statement}**
↳ Evidence: {file paths cited}
↳ If wrong: {consequence}
### {Area Name 2}
...
[If external research was done:]
### External Research Applied
- {Topic}: {Finding} (Source: {URL})
If --auto:
- If all assumptions are Confident or Likely: log assumptions, skip to write_context.
Log:
[auto] All assumptions Confident/Likely - proceeding to context capture. - If any assumptions are Unclear: log a warning, auto-select recommended alternative for
each Unclear item. Log:
[auto] {N} Unclear assumptions auto-resolved with recommended defaults.Proceed to write_context.
Otherwise: Use AskUserQuestion:
- header: "Assumptions"
- question: "These all look right?"
- options:
- "Yes, proceed" - Write CONTEXT.md with these assumptions as decisions
- "Let me correct some" - Select which assumptions to change
If "Yes, proceed": Skip to write_context. If "Let me correct some": Continue to correct_assumptions.
The assumptions are already displayed above from present_assumptions.Present a multiSelect where each option's label is the assumption statement and description is the "If wrong" consequence:
Use AskUserQuestion (multiSelect):
- header: "Corrections"
- question: "Which assumptions need correcting?"
- options: [one per assumption, label = assumption statement, description = "If wrong: {consequence}"]
For each selected correction, ask ONE focused question:
Use AskUserQuestion:
- header: "{Area Name}"
- question: "What should we do instead for: {assumption statement}?"
- options: [2-3 concrete alternatives describing user-visible outcomes, recommended option first]
Record each correction:
- Original assumption
- User's chosen alternative
- Reason (if provided via "Other" free text)
After all corrections processed, continue to write_context with updated assumptions.
Auto mode: Should not reach this step (--auto skips from present_assumptions).
Create phase directory if needed. Write CONTEXT.md using the standard 6-section format.File: ${phase_dir}/${padded_phase}-CONTEXT.md
Map assumptions to CONTEXT.md sections:
- Assumptions →
<decisions>(each assumption becomes a locked decision: D-01, D-02, etc.) - Corrections → override the original assumption in
<decisions> - Areas where all assumptions were Confident → marked as locked decisions
- Areas with corrections → include user's chosen alternative as the decision
- Folded todos → included in
<decisions>under "### Folded Todos"
# Phase {PHASE}: {phase_name} - Context
**Gathered:** {date} (assumptions mode)
**Status:** Ready for planning
<domain>
## Phase Boundary
{Domain boundary from ROADMAP.md - clear statement of scope anchor}
</domain>
<decisions>
## Implementation Decisions
### {Area Name 1}
- **D-01:** {Decision - from assumption or correction}
- **D-02:** {Decision}
### {Area Name 2}
- **D-03:** {Decision}
### the agent's Discretion
{Any assumptions where the user confirmed "you decide" or left as-is with Likely confidence}
### Folded Todos
{If any todos were folded into scope}
</decisions>
<canonical_refs>
## Canonical References
**Downstream agents MUST read these before planning or implementing.**
{Accumulated canonical refs from analyze step - full relative paths}
[If no external specs: "No external specs - requirements fully captured in decisions above"]
</canonical_refs>
<code_context>
## Existing Code Insights
### Reusable Assets
{From codebase scout + Explore subagent findings}
### Established Patterns
{Patterns that constrain/enable this phase}
### Integration Points
{Where new code connects to existing system}
</code_context>
<specifics>
## Specific Ideas
{Any particular references from corrections or user input}
[If none: "No specific requirements - open to standard approaches"]
</specifics>
<deferred>
## Deferred Ideas
{Ideas mentioned during corrections that are out of scope}
### Reviewed Todos (not folded)
{Todos reviewed but not folded - with reason}
[If none: "None - analysis stayed within phase scope"]
</deferred>
Write file.
Write audit trail of assumptions and corrections.File: ${phase_dir}/${padded_phase}-DISCUSSION-LOG.md
# Phase {PHASE}: {phase_name} - Discussion Log (Assumptions Mode)
> **Audit trail only.** Do not use as input to planning, research, or execution agents.
> Decisions captured in CONTEXT.md - this log preserves the analysis.
**Date:** {ISO date}
**Phase:** {padded_phase}-{phase_name}
**Mode:** assumptions
**Areas analyzed:** {comma-separated area names}
## Assumptions Presented
### {Area Name}
| Assumption | Confidence | Evidence |
| ----------- | -------------------------- | ------------ |
| {Statement} | {Confident/Likely/Unclear} | {file paths} |
{Repeat for each area}
## Corrections Made
{If corrections were made:}
### {Area Name}
- **Original assumption:** {what the agent assumed}
- **User correction:** {what the user chose instead}
- **Reason:** {user's rationale, if provided}
{If no corrections: "No corrections - all assumptions confirmed."}
## Auto-Resolved
{If --auto and Unclear items existed:}
- {Assumption}: auto-selected {recommended option}
{If not applicable: omit this section}
## External Research
{If research was performed:}
- {Topic}: {Finding} (Source: {URL})
{If no research: omit this section}
Write file.
Commit phase context and discussion log:pi-gsd-tools commit "docs(${padded_phase}): capture phase context (assumptions mode)" --files "${phase_dir}/${padded_phase}-CONTEXT.md" "${phase_dir}/${padded_phase}-DISCUSSION-LOG.md"
Confirm: "Committed: docs(${padded_phase}): capture phase context (assumptions mode)"
Update STATE.md with session info:pi-gsd-tools state record-session \
--stopped-at "Phase ${PHASE} context gathered (assumptions mode)" \
--resume-file "${phase_dir}/${padded_phase}-CONTEXT.md"
Commit STATE.md:
pi-gsd-tools commit "docs(state): record phase ${PHASE} context session" --files .planning/STATE.md
Created: .planning/phases/${PADDED_PHASE}-${SLUG}/${PADDED_PHASE}-CONTEXT.md
## Decisions Captured (Assumptions Mode)
### {Area Name}
- {Key decision} (from assumption / corrected)
{Repeat per area}
[If corrections were made:]
## Corrections Applied
- {Area}: {original} → {corrected}
[If deferred ideas exist:]
## Noted for Later
- {Deferred idea} - future phase
---
## ▶ Next Up
**Phase ${PHASE}: {phase_name}** - {Goal from ROADMAP.md}
`/gsd-plan-phase ${PHASE}`
<sub>`/new` first → fresh context window</sub>
---
**Also available:**
- `/gsd-plan-phase ${PHASE} --skip-research` - plan without research
- `/gsd-ui-phase ${PHASE}` - generate UI design contract (if frontend work)
- Review/edit CONTEXT.md before continuing
---
- Parse
--autoflag from $ARGUMENTS - Sync chain flag:
if [[ ! "$ARGUMENTS" =~ --auto ]]; then pi-gsd-tools config-set workflow._auto_chain_active false 2>/dev/null fi - Read chain flag and user preference:
AUTO_CHAIN=$(pi-gsd-tools config-get workflow._auto_chain_active 2>/dev/null || echo "false") AUTO_CFG=$(pi-gsd-tools config-get workflow.auto_advance 2>/dev/null || echo "false")
If --auto flag present AND AUTO_CHAIN is not true:
pi-gsd-tools config-set workflow._auto_chain_active true
If --auto flag present OR AUTO_CHAIN is true OR AUTO_CFG is true:
Display banner:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
GSD ► AUTO-ADVANCING TO PLAN
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Context captured (assumptions mode). Launching plan-phase...
Launch: Skill(skill="gsd-plan-phase", args="${PHASE} --auto")
Handle return: PHASE COMPLETE / PLANNING COMPLETE / INCONCLUSIVE / GAPS FOUND (identical handling to discuss-phase.md auto_advance step)
If neither --auto nor config enabled:
Route to confirm_creation step.
<success_criteria>
- Phase validated against roadmap
- Prior context loaded (no re-asking decided questions)
- Codebase deeply analyzed via Explore subagent (5-15 files read)
- Assumptions surfaced with evidence and confidence levels
- User confirmed or corrected assumptions (~2-4 interactions max)
- Scope creep redirected to deferred ideas
- CONTEXT.md captures actual decisions (identical format to discuss mode)
- CONTEXT.md includes canonical_refs with full file paths (MANDATORY)
- CONTEXT.md includes code_context from codebase analysis
- DISCUSSION-LOG.md records assumptions and corrections as audit trail
- STATE.md updated with session info
- User knows next steps </success_criteria>