context-discipline

Context discipline (Avacchedaka + Sublation + Witness)

This skill teaches the agent the three operations that keep long-context sessions epistemically clean: typed insertion, sublation on conflict, and boundary-triggered compaction. It is automatically activated in any session where the Pratyakṣa Context Engineering Harness is installed.

When to use

Activate this skill the moment any of these signals appear:

• the conversation has accumulated more than ~30k tokens
• the agent is about to call context_retrieve or context_window and feed the result to a planning step
• a fresh tool output contradicts something the agent stated earlier
• the user says "what's the current state of X?", "summarise our progress", or asks the agent to pick up a paused thread
• the agent is about to compact / truncate the session and would otherwise drop important typed claims indiscriminately

If none of these apply, do nothing — the harness is opt-in by design.

Three operations (in order)

1. Insert with type, not just text

Every claim that the agent might later retrieve goes through context_insert with an Avacchedaka triple (qualificand, qualifier, condition):

Field	Meaning	Example
qualificand	What is being qualified — the subject of the claim	"redis_default_port"
qualifier	The qualifier or property being asserted	"port_number"
condition	AND-conjunctive limitor tokens scoping the claim	"version=7.2 AND env=prod"
precision	Calibrated confidence in [0, 1]	0.95 (from official docs)
provenance	Source URL, file:line, or tool call id	"https://redis.io/docs/..."

Without the triple, two claims with the same surface text but different contexts collapse into one — exactly the silent corruption we want to prevent. Two redis_default_port claims with different version= conditions are first-class distinct.

2. Sublate when newer evidence contradicts older claims

When the agent finds a higher-precision claim that contradicts an older one, do not delete the older claim. Call sublate_with_evidence instead:

sublate_with_evidence({
  older_id: "claim_42",
  newer_content: "Redis 7.2 changed the default to 0.0.0.0:6379",
  newer_precision: 0.97,
  qualificand: "redis_default_port",
  qualifier: "port_number",
  condition: "version=7.2 AND env=prod",
  provenance: "https://redis.io/docs/latest/operate/oss_and_stack/management/config/"
})

Sublation preserves the older claim with precision=0.0 and sublated_by=newer_id — auditable, but invisible to future context_retrieve calls. This is the Bādha (sublation) operation from classical Vedānta: the rope is not erased when it's seen-as-rope-not-snake; the snake-perception is sublated by the rope-perception.

3. Compact at event boundaries, not on token-count alone

When the agent detects a discourse boundary — task switch, phase transition, user interrupt with a fresh topic — call boundary_compact on the recent text window. The harness's surprise-spike detector identifies the boundary and triggers a scoped compact so only low-precision claims in the closing phase are dropped. Never call generic "summarize and truncate" on a long session that has typed claims in the store; you will silently lose auditability.

Witness invariants (Sākṣī)

Any invariant that must remain true across every model call this session — the user's identity, the project's hard rules, the active hypothesis under test — should be set once via set_sakshi. The witness-prefix skill guarantees it is pushed as a real Claude system field at every model call (NOT inlined into user content, which is what naive prompt-stitching implementations do — see G9 in the project's docs/v0_retrospective.md).

Anti-patterns the harness catches

Anti-pattern	Harness response
Inserting two contradictory claims with the same id	context_insert rejects unless overwrite=true
Dropping older context to "fit the window"	compact requires a scope; boundary_compact requires a real boundary
Treating sakshi as user content	set_sakshi returns a system_message block to push verbatim
Calling context_retrieve without a qualificand	The schema requires it — no implicit cross-topic drift

Diagnostic commands

• /context-status — show per-qualificand element counts and mean precision
• /sublate <older_id> — open the sublation dialog for a specific element
• /compact-now — force boundary compaction on the recent window
• /budget — show tokens used so far in this session