Resume work from a previous session checkpoint. Uses `resume-context` CLI to load feature, checkpoint, git state, and session history in a single call, then synthesizes a focused resumption summary before continuing. Use at the start of a new session to restore context from a previous checkpoint.
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Resume work from a previous session checkpoint. Uses `resume-context` CLI to load feature, checkpoint, git state, and session history in a single call, then synthesizes a focused resumption summary before continuing. Use at the start of a new session to restore context from a previous checkpoint.
Where $DISCOVER is the absolute path to scripts/discover.sh within this skill's directory.
Path safety — shell state does not persist between tool calls, so you must provide full script paths on each call:
Use $HOME instead of the literal home directory (e.g., bash "$HOME/code/…/discover.sh", not bash "/Users/name/…/discover.sh"). This prevents username hallucination.
Copy values from tool output. When reusing a value returned by a previous command (like $PROJECT_ROOT), copy it verbatim from that command's output. Never retype a path from memory.
: if a script call fails with "No such file", the path is wrong — STOP and re-derive from the skill-loading context.
Verify on first call
Never ignore a non-zero exit. If any script in this skill fails, stop and report the error before continuing.
Store the output as $PROJECT_ROOT. If the command fails, inform the user and stop.
Where $VALIDATE is the absolute path to scripts/validate.sh within this skill's directory. Apply the path safety rules from Step 0 ($HOME, copy from output). Outputs $FEATURE_NAME on success; on failure, STOP immediately — do not continue with an unvalidated path.
Set $FEATURE_DIR to $PROJECT_ROOT/.dev/$FEATURE_NAME.
Step 2: Load Context via resume-context
Make a single call to resume-context — this replaces the previous Steps 2–5 (git-state, checkpoint-read, feature-show, validity comparison):
Where $CLI is the absolute path to scripts/dev-workflow.cjs within this skill's directory. Apply the path safety rules from Step 0 ($HOME, copy from output).
Parse the JSON output. It contains all the data previously gathered across multiple tool calls:
currentPhasePrd: Extracted markdown section for the current phase only (not the full master plan)
referenceFiles: File paths from the master plan's "Reference Files" section
sessionDigest: Distilled digest of the older session tail — { sessionCount, consolidatedThrough, generated, aggregate, decisions[] } — written by /dev-checkpoint once a feature crosses the consolidation threshold (≥10 sessions). null below the threshold or when no digest has been written yet.
recentSessionHistory: The recent raw session window (last N, default 5; --sessions=all for all). Older sessions beyond the window are summarized in sessionDigest when present. (Replaces the former sessionHistory field.)
accumulatedDecisions: Bounded decision set — when a sessionDigest exists, it is the digest's carried-forward decisions ∪ the recent window's decisions (deduplicated); without a digest, the full deduplicated union across all sessions (legacy behavior).
The LLM no longer calls: git-state.sh, checkpoint-read --json, feature-show --json. It no longer compares branches manually. It no longer reads the full master plan (currentPhasePrd contains just the relevant section).
Check the exit code. A non-zero exit code means the command failed (e.g., feature directory not found, no checkpoint). Never ignore a non-zero exit — if resume-context fails, stop and report the error. Do not attempt to reconstruct the data from separate tool calls.
Step 3: Check Context Validity
Read the validity field from the resume-context output:
fresh: Branch matches, checkpoint is recent (< 3 days old). Proceed normally.
stale: Branch matches but checkpoint is old. Show an informational warning, then proceed.
drifted: Branch mismatch. Warn and ask: "Checkpoint was on X, you're on Y. Switch or continue?" Wait for user response before proceeding.
If "drifted", the validityDetails field shows both branch names for the prompt.
Step 4: Present Resumption Summary
Synthesize the resumption summary from the resume-context output. This is the irreplaceable LLM step — the CLI handles the mechanics; the LLM handles the judgment.
**Status**: [feature.currentPhase.title] — [feature.progress.done]/[feature.progress.total] ([feature.progress.percent]%)
**Last session**: [Derive from checkpoint.context field]
**Decisions**: [checkpoint.decisions, or "None recorded"]
**Session history**: [If sessionDigest present: "sessions 1–{sessionDigest.consolidatedThrough} consolidated into a digest + {recentSessionHistory.length} recent raw"; otherwise "{recentSessionHistory.length} sessions tracked"] — {accumulatedDecisions.length} decisions carried forward
**Feature arc**: [Only when sessionDigest present — summarize sessionDigest.aggregate so the distilled older-session history isn't lost. Omit this line entirely when sessionDigest is null.]
**Watch out for**: [checkpoint.blockers, or "Nothing flagged"]
**Current phase PRD**: [currentPhasePrd — summarize or reference; don't re-read the full master plan]
**Start with**: [First concrete action from checkpoint.nextAction field]
Incorporate accumulatedDecisions when relevant — decisions from earlier sessions may contextualize the current step. When a sessionDigest is present, lean on its aggregate for the older history rather than the raw recentSessionHistory window, which only covers the last few sessions.
Wait for go-ahead — do not proceed until the user confirms.
Step 5: Handling Discrepancies
Situation
Action
File differs from checkpoint
Proceed, note drift
Key file missing or renamed
STOP — ask how to proceed
New files not in checkpoint
Proceed, mention them
PRD files missing
STOP — cannot resume without PRD
Step 6: Read Key Files and Reference Patterns
Before beginning work:
The currentPhasePrd field from resume-context contains the relevant phase section — read it instead of the full master plan
Use referenceFiles from the resume-context output to identify which files to read — these are the PRD's recommended reference implementations
Read 2-3 of those reference files and match their conventions (naming, structure, APIs, error handling) in new code
Optionally, check $HOME/.dev-wiki/index.md for cross-project context (related features, prior art). Skip silently if absent.
Never write new code from scratch when similar code already exists in the codebase.
Step 7: Begin Work
After confirmation, proceed with the first action from the agent's summary. Follow the PRD phases and gates.
CRITICAL: PHASE GATE ENFORCEMENT
After completing a phase's work, run the CLI to mechanically verify gate status:
node "$CLI" gate-check --json --dir"$FEATURE_DIR"
Where $CLI is the absolute path to scripts/dev-workflow.cjs within this skill's directory.
The command always returns exit 0 on success (exit 1 on error). Check the atGate and allComplete fields in the JSON output to determine gate status. Use this to confirm phase completion rather than relying on visual inspection of markdown markers.
At every gate (whether detected by gate-check or by ⏸️ **GATE**: markers in the PRD) — this is a HARD STOP:
STOP — Do not proceed to the next phase
Report what was accomplished
Ask: "Phase [N] complete. Continue to Phase [N+1] or /dev-checkpoint?"
Wait for explicit user response before continuing
If gate-check returned allComplete: true — or the next phase is the final verification/testing phase — extend the gate question with: "This was the last implementation phase. Consider running /dev-review for a PRD-vs-implementation alignment report before final testing."
STEP-LEVEL STOPS
After completing each implementation step within a phase:
Report what was completed
Ask: "Step done. Continue to next step?"
Wait for confirmation before proceeding
This prevents jumping ahead to the next task before the current one is tested.
PRIVACY RULES
Warn the user if checkpoint/PRD files contain: absolute paths with usernames, secrets/credentials, or personal information.
