Universal Read Output Contracts

Tracker references: pm-15yuly, pm-g8oh0f, pm-hb7ug8, pm-cxr0jb, pm-hid9g1, pm-h8tpeh, pm-5t33or, pm-wv47pf, pm-srns, pm-sb0tns, pm-gjjurs, pm-eugaqy, pm-jt8aa2, pm-kyjdne, pm-8nev0o, pm-oahhyc, pm-q4isdq, pm-e5gl05, pm-cha95z, pm-2qvq7a, and pm-xam9bt.

Agent Quick Context

Every built-in read surface uses four output dimensions: what to include, how much to return, how much the result may cost, and how to encode it. The same canonical controls work through the CLI, SDK, MCP, generated schemas, runtime contracts, and shell completions.

Dimension	CLI	SDK and MCP	Meaning
Include	--output-include <csv>	outputInclude	Retain named fields or sections, or select a declared projection mode.
Amount	--output-limit <n|unbounded>	outputLimit	Bound shared row collections.
Cost	--output-budget <n|unbounded>	outputBudget	Fail closed when even the compact result cannot fit.
Encoding	--output-format <toon|json>	outputFormat	Select the CLI renderer and record the requested encoding.

The contract covers list, context, search, get, next, health, deps, graph, history, activity, validate, events, contracts, comments, notes, files, docs, stats, and aggregate, including list aliases and ctx.

Primary row shaping follows each envelope's row_contract.row_keys, including dot-delimited arrays and object maps such as graph.nodes. Include, amount, and repeat suppression therefore remain bound to the rows the command says it returns. A result may additionally declare continuation_row_keys when a nested evidence collection must resume independently without redefining the primary rows that --output-limit bounds. Cost compaction may inspect both primary and nested collections; it does not rely on command-specific keys. The runtime uses that declaration internally on every read but omits the repeated metadata from results by default. Request --output-row-contract / outputRowContract: true when a consumer needs the row paths, jq selector, and active TOON encoding contract.

Include Modes

--output-include accepts two kinds of token. A field selector names a row field or section and narrows the computed result. A projection mode names a whole declared projection and is the canonical spelling of a command-local mode flag: brief, compact, full, summary, and counts, depending on the surface. Controls that change execution rather than projection remain separate: for example, deps --collapse <none|repeated> retains dependency-grouping semantics, and health --check-only retains refresh-suppression semantics.

Mode tokens are resolved before the command computes its rows, because a mode selects which fields exist rather than which of the computed fields survive. pm list --output-include brief is therefore exactly pm list --brief, and the two are byte-identical apart from the read_output receipt that records which spelling was used. Modes and field selectors compose: the mode selects the projection, the remaining selectors narrow it.

pm list --status open --output-include brief          # same result as --brief
pm contracts --output-include full                    # same result as --full
pm list --status open --output-include brief,id       # brief projection, id only

Read readOutputIncludeModeOptions(command) from the SDK for the exact replacement modes a surface declares. Every compatibility alias also declares semantics: "replacement" | "behavior_preserving"; generators therefore do not have to infer obligation strength from prose. The executable migration test derives all 22 projection-mode replacements from this table, invokes both spellings in a temporary tracker, and compares their useful result after removing spelling receipts and volatile run metadata. A selector that matches neither a declared mode nor any field on any returned row is refused with the legal domain, rather than returning rows with every field removed.

Runtime discovery publishes the same exact vocabulary before invocation. pm contracts --command <command> --summary returns output_projection_contracts.commands, while output_projection_contracts.global_ladder_scope: "union_not_per_command" explicitly marks output_policy.degradation_ladder as a cross-command union, not a promise that every member is accepted by the selected command. For example, list declares brief, compact, and full; health declares brief, full, and summary. The full read_output_dimensions matrix also includes projection_modes for every canonical read surface, including an explicit empty array when a surface has no whole-result modes.

Single-item detail fallback

Tracked by pm-8eclqf. When a standard get exceeds its budget, the SDK first tries the declared brief projection. It retains complete item metadata and dependencies while omitting body, linked artifacts, claim history, and schedule details. The receipt records applied_depth: brief and degradation_reason: output_budget_reached; the omission receipt names the removed sections and how to restore them. Explicit field selections, full/deep reads, trees, and continuation requests keep their selected semantics. If brief still exceeds the ceiling, normal compaction and eventual refusal apply. estimated_tokens measures the emitted envelope; omitted_result_estimated_tokens measures the useful result that could not fit.

Cross-Call Context Sessions

--output-session <json> / outputSession composes the four per-call dimensions across a request group. The caller supplies versioned state and passes the returned read_session.next_state to the next read:

{
  "version": 1,
  "id": "orientation",
  "token_budget": 4000,
  "spent_tokens": 0,
  "seen_item_ids": []
}

The session ceiling and an explicit --output-budget both bind; the smaller remaining allowance wins. Rows for item facts already present in the caller's context become { "id": "pm-a1b2", "context_ref": "session:orientation:pm-a1b2" } instead of repeating prose. References retain stable item identity and can be restored with pm get <item-id> --brief when the prior context is unavailable. The receipt reports estimated and charged tokens separately when the remaining group allowance is smaller than the minimum control envelope, plus the accumulated spend, remaining capacity, newly served items, and suppressed repeats.

The carried served-item set accepts at most 10,000 identities. A receipt always preserves identities already present in the supplied state and retains newly served identities in deterministic order until that capacity is full. When a single read crosses the boundary, seen_item_overflow_count reports how many new identities were not carried forward; those facts remain in the current envelope and may be served in full again on a later read. The emitted next_state therefore always remains valid input to the next call without silently widening the safety bound.

Session state is deliberately caller-carried: CLI processes, SDK clients, MCP hosts, and packages share the same deterministic primitive without a hidden daemon or mutable cache. Validation rejects unknown fields, invalid identifiers, unsupported schema versions, unsafe integers, and spend beyond the declared ceiling before a read executes.

The mandatory orientation calibration runs context, list, search, get, and next against both a two-item tracker and a 2,243-item tracker. Its cross-call ceilings are strict: complete serialized bytes and cumulative spend may only shrink, while repeat suppression may only hold or improve. The gate also fixes the expected unique-fact shape:

Tracker tier	Group spend / budget	Seen items	Suppressed repeats	Delivered bytes
2 items	3,820 / 20,000	2	3	15,274
2,243 items	10,156 / 20,000	106	7	40,614

These are deterministic synthetic-corpus measurements from scripts/release/context-intent-calibration.json; they contain no hosted tracker content.

Precedence and Compatibility

Resolution is deterministic: canonical controls win over command-local compatibility options, which win over intent defaults, which win over command defaults. Existing options such as --fields, --limit, --token-budget, --format, --brief, and --full remain accepted. Contract output marks them as hidden compatibility aliases and supplies a migration hint; traversal, cursor, side-effect, and streaming controls instead receive an explicit behavior-preservation hint because a static output control cannot replace their semantics. Callers that omit the four shaping dimensions retain the established data projection; the one intentional envelope correction is that repeated row_contract metadata is now opt-in.

Completeness spellings preserve their established promise. pm list-all and list-family --no-truncate imply an unbounded cost dimension when the caller does not supply --output-budget; they cannot silently return a budget-trimmed subset while claiming to return every matched row. An explicit canonical budget still has precedence and may request a bounded result deliberately. The same SDK-native rule applies to every declared complete-result spelling, including contracts --full: explicit completeness defeats only the implicit default ceiling. If an explicit caller budget cannot retain any useful domain result, the CLI still prints the parseable omission receipt but exits 2; a consumer therefore cannot interpret the missing command or action arrays as a successful empty contract.

pm list --status open --output-include id,title,status --output-limit 10
pm context --for orient --output-budget 900 --output-format toon
pm search "runtime contracts" --output-limit 5 --output-format json
pm stats --output-row-contract
pm contracts --full --json

Canonical amount, encoding, and cost controls that leave an already-fitting result unchanged do not add a read_output receipt. Explicit include projections, sessions, continuations, and --output-row-contract retain the detailed receipt. Every shaped or budget-compacted result carries a read_output receipt with the requested dimensions, precedence, observed option provenance, deterministic estimated token count, string/row compaction signals, and budget outcome. When canonical controls are present, canonical_options_used records spellings such as --output-include; the optional field is omitted when no canonical control was supplied so minimum-budget session receipts remain compact. legacy_aliases_used and migration_hints record only compatibility flags the caller actually supplied. A command-local flag synthesized internally to execute a canonical projection mode is therefore never misreported as deprecated syntax. When both spellings are supplied, both provenance arrays report that mixed invocation. Budget degradation discovers nested arrays as well as declared result rows, so validation diagnostics and other governance payloads compact their inner findings before the useful result is omitted. compacted_row_paths names every reduced collection without redefining those nested arrays as ordinary pagination rows.

The estimate is ceil(utf8_bytes / 4) for the complete envelope. The CLI uses its selected built-in JSON or TOON representation, including indentation and the final newline. SDK calls that select outputFormat use that representation; structured SDK calls without a renderer retain compact JSON measurement. Those unformatted calls keep the conservative workload-class ceiling (default_max_estimated_tokens, also used as the generated TOON ceiling). Selecting JSON explicitly opts into its larger declared rendering allowance; serializing an SDK object for measurement does not implicitly select that policy. Per-call and session receipts include their own cost and continuation disclosure. Compaction reserves room for that disclosure before returning a useful page, so an envelope cannot claim to fit merely because its rows fit before receipts were added. Each selected collection uses a measured prefix search instead of irreversible halving, retaining the largest prefix that fits with the other collections held fixed. applied_limit and both continuation forms follow the retained item count. A complete page and a truncated page have different mandatory metadata costs; row counts across that boundary are not an unconditional monotonicity guarantee. An infeasible ceiling returns the explicit omission result.

For context and next, that final projection is also the usage-feedback delivery boundary. The SDK carries a JSON-invisible serve receipt through intent and universal output projection, then appends the final omitted decision and exact emitted item ids. An infeasible output budget records zero delivered rows; it can never leave the packer's earlier inclusions marked as served. PmClient finalizes the receipt before resolving, while the CLI finalizes it from the exact active result passed to telemetry and after-command hooks.

When rows are dropped to satisfy a ceiling, the result also carries output_budget_truncation, naming the binding budget and its source, any explicitly requested dimension the budget overrode, every compacted collection path, and executable CLI/SDK/MCP recovery options — a default ceiling can override an explicit --output-limit unbounded, and that override is reported rather than silent. If a producer supplied an opaque item-page cursor, compaction rebases it to the last row actually returned and reports continuation_cursor_rebased: true; following the cursor therefore cannot skip rows removed from the middle of a producer page. If no useful content can fit, PmReadOutputBudgetExceeded provides a discriminated omission result, a compact { outputBudget: "unbounded" } recovery object, and omitted_result_estimated_tokens, the last useful-result estimate before omission; use isReadOutputBudgetExceeded before accessing result-specific fields. The encoding control --output-format <toon|json> also applies to mutation envelopes, including the mutation modes of comments, notes, files, and docs. Other universal read controls remain rejected before mutation execution. --json remains supported.

Cursor recovery carries the opaque value once in recovery.cursor and declares the accepting cli, sdk, and mcp binding beside it. This avoids serializing the same cursor once per transport while retaining an executable, machine-readable binding for each surface. When restore_with is "recovery", consumers must resolve the adjacent recovery object: pass its cursor value through cli (--output-cursor) or the sdk/mcp field (outputCursor) while repeating the original query. The string is a reference, not a shell command. Other restore_with values provide retry guidance when no declared row collection can continue.

Budget-compacted declared row paths are independently resumable. The disclosure's continuations entries name the row path, retained/remaining/total counts, and an opaque cursor. continuation_kind distinguishes a rebased producer cursor, a universal output cursor, and a terminal page; next_cursor mirrors the first universal entry for ordinary one-path consumers. Replay the same query and budget with --output-cursor <cursor> / outputCursor. The cursor validates the command, declared continuation path, total, and canonical content fingerprint before slicing, so same-cardinality content changes and other stale replays fail closed instead of skipping evidence. A bounded recovery therefore does not require replacing a 600-token request with an unbounded multi-megabyte response; recovery_budget_multiplier: 1 declares that each next page retains the same useful-result ceiling.

Health continuation fingerprints use a versioned nonvolatile-snapshot policy for the checks row. Command-lifecycle telemetry may refresh timestamps, queue counters, and pending spans between pages, so the policy excludes only those explicit direct fields from the telemetry row's details. The row identity, verdict, enabled state, capture level, endpoint, environment overrides, probe result, and failure errors remain fingerprinted. Every other health check also remains fully fingerprinted; changing storage, integrity, history, or another stable check fails closed. PM_READ_OUTPUT_CONTINUATION_FINGERPRINT_POLICIES publishes the exact ignored detail fields per dynamic row. Other commands still fingerprint every field, including fields with the same spelling, and any policy-version change fails closed against an older cursor.

When compaction affects a nested collection that is not declared resumable, the recovery budget is derived from the larger of the binding ceiling plus one and the measured pre-compaction result. The SDK adds a 25% envelope margin and rounds up to 100 tokens. The CLI, SDK, and MCP bindings therefore receive the same finite value, which is strictly larger than the request that already truncated. If safe-integer arithmetic cannot represent that value, recovery uses unbounded explicitly. resolveReadOutputRecoveryBudget() exposes the versioned pure calculation to package authors; there is no hard-coded smaller retry and no recovery cycle.

Assurance declares budget_retention_policy: verdict_priority: failing block, warn, and observe rows precede retired and passing rows while preserving order within each class. assertions_total remains the pre-projection denominator, and the assertions row path uses the same continuation primitive. A blocking verdict consequently keeps its causal evidence on the first bounded page even when the failing assertion was declared last.

Diagnostic gates retain bounded, actionable predicates. Every metadata field required by the active validation profile emits a numeric count, including zero, so assurance expressions can distinguish a clean check from a missing key. Health retains at most 100 warning rows and reports warning_count, warning_limit, and warnings_truncated; --strict-exit defaults to summary projection unless --full is explicit, keeping the failing check identities and warning codes inside the ordinary budget.

Failure transports use the companion Diagnostic Output Contracts. Those contracts bind errors, warnings, validation summaries, and recovery bundles independently from successful command results. Corrective actions precede explanation and survive every degradation stage, while untruncated JSON avoids a per-call receipt tax.

SDK and Package Usage

Typed PmClient read methods accept PmReadOutputOptions directly:

import { PmClient } from "@unbrained/pm-cli/sdk";

const pm = new PmClient({ pmRoot: ".agents/pm" });
const result = await pm.list({
  status: "open,in_progress",
  outputInclude: "id,title,status",
  outputLimit: 10,
  outputBudget: 800,
  outputRowContract: true,
  outputSession: {
    version: 1,
    id: "orientation",
    token_budget: 4000,
    spent_tokens: 0,
    seen_item_ids: [],
  },
  // Supply outputCursor from output_budget_truncation to fetch the next row page.
});

Package authors should use the exported read-output contracts instead of inventing package-local limit or projection vocabularies. PM_READ_OUTPUT_SURFACE_CONTRACTS is the machine-readable matrix, PM_READ_OUTPUT_CONTINUATION_FINGERPRINT_POLICIES declares stable-snapshot exceptions, and resolveReadOutputDimensions plus applyReadOutputDimensions are the shared execution primitives.

Drift Gates

The full runtime contract reports every surface and all four dimensions. Strict SDK/MCP schemas expose the canonical camelCase keys, and the generated contract fixture catches surface or schema drift:

pnpm contracts:check
node scripts/run-tests.mjs test -- tests/unit/sdk/read-output-contracts.spec.ts
node scripts/run-tests.mjs test -- tests/unit/sdk/read-output-migration-hints.spec.ts
node scripts/run-tests.mjs test -- tests/unit/commands/workspace/completion-command.spec.ts

Ein Problem melden